Diversity Polynomials for the Analysis of Temporal Correlations in Wireless Networks
The interference in wireless networks is temporally correlated, since the node or user locations are correlated over time and the interfering transmitters are a subset of these nodes. For a wireless network where (potential) interferers form a Poisson point process and use ALOHA for channel access, we calculate the joint success and outage probabilities of transmissions over a reference link. The results are based on the diversity polynomial, which captures the temporal interference correlation. The joint outage probability is used to determine the diversity gain (as the SIR goes to infinity), and it turns out that there is no diversity gain in simple retransmission schemes, even with independent Rayleigh fading over all links. We also determine the complete joint SIR distribution for two transmissions and the distribution of the local delay, which is the time until a repeated transmission over the reference link succeeds.
1.1Motivation and contributions
The locations of interfering transmitters in a wireless network are static or subject to a finite level of mobility. As a result, the interference power is temporally correlated, even if the transmitters are chosen independently at random from the total set of nodes in each slot. The interference correlation has been largely ignored until recently, although it can have a drastic effect on the network performance. In this paper, we provide a comprehensive analysis of the joint success and outage probabilities of multiple transmissions over a reference link in a Poisson network, where the potential interferers form a static Poisson point process (PPP) and the actual (active) interferers in each time slot are chosen by an ALOHA multiple-access control (MAC) scheme. The results show that for some network parameters, ignoring interference correlation leads to significant errors in the throughput and delay performance of the link under consideration.
The Poisson network model has served as an important base-line model for ad hoc and sensor networks for several decades and later also for mesh and cognitive networks. More recently, it has also been gaining relevance for cellular systems, where base stations are increasingly irregularly deployed, in particular in heterogeneous networks . Consequently, the results in this paper may find applications in a variety of networks.
The paper makes four contributions:
We introduce the diversity polynomial and provide a closed-form expression for the joint success probability of transmissions in a Poisson field of interferers with independent Rayleigh fading and ALOHA channel access (Section III).
We show that there is no temporal diversity gain (due to retransmission), irrespective of the number of retransmissions—in stark contrast to the case of independent interference (Section III.D).
We provide the complete joint SIR distribution for the case of two transmissions and show that the probability of succeeding at least once is minimized if the two transmissions occur at the same rate (Section IV).
We determine the complete distribution of the local delay, which is the time it takes for a node to transmit a packet to a neighboring node if a failed transmission is repeated until it succeeds (Section V).
The first paper explicitly addressing the interference correlation in wireless networks is , where the spatio-temporal correlation coefficient of the interference in a Poisson network is calculated. It is also shown that transmission success events and outage events are positively correlated, but their joint probability is not explicitly calculated. In , the temporal interference correlation coefficient is determined for more general network models, including the cases of static and random node locations that are known or unknown, channels without fading and fading with long coherence times, and different traffic models. In , the loss in diversity is established for a multi-antenna receiver in a Poisson field of interference. The probability that the SIR at antennas jointly exceeds a threshold is determined in closed form. This result is a special case of the main result in this paper, where the focus is on temporal correlation. More recently,  studied the benefits of cooperative relaying in correlated interference, for both selection combining and maximum ratio combining (MRC), while  analyzed on the impact correlated interference has on the performance of MRC at multi-antenna receivers.
A separate line of work focuses on the local delay, which is the time it takes for a node to connect to a nearby neighbor. The local delay, introduced in  and further investigated in , is a sensitive indicator of correlations in the network. In  the two lines of work are combined and approximate joint temporal statistics of the interference are used to derive throughput and local delay results in the high-reliability regime. In , the mean local delay for ALOHA and frequency-hopping multiple access (FHMA) are compared, and it is shown that FHMA has comparable performance in the mean delay but is significantly more efficient than ALOHA in terms of the delay variance.
We consider a link in a Poisson field of interferers, where the (potential) interferers form a uniform Poisson point process (PPP) of intensity . The receiver under consideration is located at the origin , and it attempts to receive messages from a desired transmitter at location , where , which is not part of the PPP. Time is slotted, and the transmission over the link from to is subject to interference from the nodes in , which use ALOHA with transmit probability . The desired transmitter is transmitting in each time slot. The transmit power level at all nodes is fixed to , and the channels between all node pairs are subject to power-law path loss with exponent and independent (across time and space) Rayleigh fading.
The signal-to-interference ratio (SIR) at in time slot is then given by
where is the set of active interferers in time slot and is a family of independent and identically distributed (iid) exponential random variables with mean . In each time slot , forms a PPP of intensity , but the point processes and are dependent for all , since they are subsets of the same PPP . In the extreme case where , , . This dependence is what makes the following analysis non-trivial.
3The Diversity Polynomial and the Joint Success Probability
We use a standard SIR threshold model for transmission success and denote by the transmission success event in time slot . We first focus on the probabilities of the joint success events
To calculate this probability, we introduce the diversity polynomial.
The second binomial can be expressed as
The first four diversity polynomials are
For fixed and , is concave increasing from to , for .
For fixed and , is convex increasing from to , for .
See Appendix A.
The parameter is related to the spatial contention parameter introduced in . For Poisson networks, , hence .
When evaluating as a function of , it must be considered that is itself a function of , not just .
For , the result reduces to the well-known single-transmission result , for all .
If (), , which means the success events become independent. At the same time, , so .
If (), , which is the case of maximum correlation. At the same time, , which is the smallest possible value.
If and , for all , so the success events are fully correlated (despite the iid Rayleigh fading), i.e.,
and . This is a strict hard-core condition, i.e., all transmissions succeed if there is no interferer within distance .
If , the diversity polynomial simplifies to the one introduced in  for the SIMO case, where it quantifies the spatial diversity instead of the temporal diversity:
As these remarks show, the diversity polynomial characterizes the dependence between the success events and the diversity achievable with multiple transmissions.
Defining the diversity as the log-likelihood ratio of the success probability with transmissions and the success probability of one transmission, the multi-transmission scheme would achieve a diversity of
if success events were independent. The actual diversity is, however,
So, for small , the diversity loss is . An immediate important consequence of Thm. 1 is the following result for the conditional success probability of succeeding at time after having succeeded times:
Fig. ? displays the conditional success probability for . It can be seen that succeeding once or twice drastically increases the success probability if is not too small. This illustrates that treating interference as independent may result in significant errors.
3.2Alternative forms of the diversity polynomial
be the polynomial of order with roots at and . Thus equipped, we can write the diversity polynomial as
by observing that
Using the Stirling numbers of the first kind , the falling factorial
Rewriting the binomial as
This expansion in is useful since in most situations. For , the polynomial in this form is
For , since , , we have from
This expression is useful as an approximation for general if (or ).
Alternatively, can be expressed as a polynomial in as
In this last expression, the term for is . This is the polynomial in obtained when . Conversely, when , it is .
3.3Event correlation coefficients
Let be the indicator that occurs. The correlation coefficient between and , , is
The correlation coefficient for is illustrated in Fig. ? as a function of and . It reaches its maximum of at and . While it is decreasing in , it is not monotonic in at .
Since the correlation coefficient is decreasing with for all “interior” values , we have the bound
Since , we have , thus the failure events are correlated in exactly the same way as the success events: If , then .
3.4Joint and conditional outage probabilities
The dependence between two success events can be quantified by the ratio of the probabilities of the joint event to the probability of the same events if they were independent. We obtain
which is consistent with the fact that the correlation coefficient is positive. The positive correlation is also apparent from the conditional probability that the second attempt succeeds when the first one did, which is
The probability of (at least) one successful transmission in attempts follows from the inclusion-exclusion formula
For the joint outage it follows that
which is consistent with the previous observation that failure events are also positively correlated.
From , the success probability given a failure follows as
which is maximized at , where it is .
with equality if and only if . Since , is achieved by either letting the interferer density , the transmission distance , or the SIR threshold go to .
Next we examine the conditional outage probability of an outage in slot given that outages occurred in slots through . Since as , one would expect this conditional outage probability to go to zero in the limit. Interestingly, this is not the case.
This is in stark contrast to the independent case, where this limit is obviously . The actual asymptotic conditional outage probability is increasing in and reaches as .
Conversely, we have for the conditional success probability given failures
Fig. ? illustrates the conditional outage probability after failures for and .
3.5Diversity gain of retransmission scheme
This is analogous to the standard definition in noise-limited systems, where diversity is defined as the exponent of the error probability as the (mean) SNR increases to infinity, see, e.g., . In our interference-limited system, the relevant quantity is the SIR.
To calculate the diversity, we need to first establish the connection between the mean SIR and the parameter . The can be increased by either increasing the received signal power or by decreasing the interference. Either way, we find that :
If we increase the received power by increasing the transmit power at the desired transmitter, we have . Since increasing and decreasing have the same effect on the success probability , we have and thus .
If we increase the received power by reducing the link distance , we have . Since , we obtain .
If we reduce the interference by decreasing the intensity of the PPP, we have since the interference is a stable random variable with characteristic exponent . Since and , we again have .
In conclusion, letting is the same as letting , and we can express the diversity as
Next we need a lemma that establishes expansions on the probability of succeeding at least once in transmissions.
Proof: See Appendix B.
In contrast, with independent interference, the diversity gain would be
So, retransmissions in (static) Poisson networks provide no diversity gain.
Conversely, fixing and varying , we have from and the fact that
so if the SIR is increased by decreasing , full diversity is restored. The difference in the behavior lies in the fact that captures the static components of the network, while reducing reduces the dependence between the interference power in different time slots.
Alternatively, the diversity could be defined on the basis of instead of , which would yield diversity in the independent case (and diversity in reality). This value may be better aligned with the intuition of what the diversity gain should be with independent transmission attempts.
Using a Taylor expansion on the exponential terms in (valid for small or , or high success rate in general), we have
So the joint outage probability in attempts is
3.6Effect of bounded path gain
Here we derive the conditional success probability for the case where the (mean) path gain is bounded, i.e., instead of assuming a gain of for a link of distance , we employ a path gain of . Equivalently, the path loss is .
Proof. See Appendix C.
The middle term in the expression for is the one for the unbounded path gain, whereas the other two account for the difference between the unbounded and bounded case. Since , the bounded and unbounded cases coincide as , i.e., for large SIR thresholds or distances of the desired link. Even for and , the difference is insignificant, as Fig. ? illustrates. The figure replicates Fig. ? for bounded path gain and shows the same behavior: Succeeding once or twice significantly increases the success probability for not too small. This suggests that the conclusions and trends observed in the unbounded case also hold in the bounded case.
4The Two-Transmission Case with Different SIR Thresholds
Here we explore the case of but with different thresholds, i.e., we focus on the events and . This case is of interest for two reasons: First it leads directly to the complete joint SIR distribution, second it is useful to provide guidance on how the rate of transmission affects the probabilities of succeeding twice or succeeding after a failure.
Proof. See Appendix D.
Since the joint success probability is symmetric in and , the expression is even in , and it can be tightly bounded by its quadratic Taylor expansion
With independent interference, we would have . As expected,
which shows that transmission success events are positively correlated for all thresholds , .
The joint SIR distribution follows from Thm. ? as
The next result shows that is an extremal point of the joint outage probability.
Proof: See Appendix E.
Hence the probability of succeeding at least once in two transmissions can be increased by using asymmetric thresholds , corresponding to . Conversely, the joint outage probability is maximized at .
Since is an even function of , it can be expressed as
where and is the second derivative at . and are given by
Since is the global minimum, we know that .
In Fig. ?, exact curves for and the quadratic approximations are shown for and . It can be observed that the approximation is quite accurate (slightly optimistic, in fact) for , which corresponds to .
4.2Comparison with two independent transmissions
Here we investigate three cases where actual success probabilities are compared with the probabilities obtained if the two success events were independent.
Joint success probability
Since transmission success events are positively correlated, we expect that the link can accommodate a certain level of asymmetry in the thresholds for the two transmissions. To explore this, we find the value of such that
or, writing out the probabilities,
To find an approximate value of for which this holds we use . Taking logarithms and dividing by yields
and we obtain
This is the level of SIR asymmetry that can be afforded thanks to the positive correlation. The resulting joint success probability will be slightly higher than , since is a (tight) bound.
Assuming a transmission rate of nats/s/Hz for an SIR threshold of , which can be achieved if Gaussian signaling is employed, the positive correlation translates to a rate gain or throughput gain since
is increasing in . Compared to the symmetric case, the throughput gain is
Probability of succeeding at least once
Alternatively, one may want to ensure that the probability of succeeding at least once in two transmissions is the same as in the independent case. This is guaranteed if
To solve this equation for , we approximate , which is valid since is the minimum of per Cor. ? and the curvature given by in is small
is calculated in and denoted by .
In Fig. ?, the design procedure is illustrated. At , the probabilities and are shown in solid and dashed curves, respectively. First we observe that while independent transmission success would yield a success probability of at , the actual success probability is slightly less than . The two curves intersect at . So if a threshold of was used in the independent case and thresholds were used for the two transmissions in the dependent case, the success probability would be about 91% in both cases. So the penalty in the SIR threshold due to the correlation is about . This is the necessary reduction in the threshold for the second transmission to achieve the same two-transmission success probability as in the independent case.
Since the intersection between the solid and dashed curves cannot be calculated in closed form, the intersection between (the dash-dotted curve) is used instead, which yields the slightly conservative value of .
Conditional success probability after failure
Lastly, one may want to choose the threshold for the second transmission such that the conditional success probability after a failure is still as large as the success probability in the independent case, i.e., the problem is to find such that
This should be the same as . The resulting equation
can be numerically solved for .
5Random Link Distance and Local Delay
5.1Random link distance
Now we let the transmission distance be a random variable (which is constant over time), denoted by . We consider the case where is Rayleigh distributed with mean , since this is the nearest-neighbor distance distribution in a PPP of intensity . This situation models a network where the receivers form a PPP of intensity , independently of the PPP of (potential) transmitters of intensity , and each transmitter attempts to communicate to its closest receiver. To remain consistent with the assumption of the typical receiver residing at the origin and its desired transmitter being active in each time slot, we add the point to the receiver PPP and an always active transmitter at distance . The joint success probability over this link of random distance is denoted by .
Expanding the diversity polynomial, can be written for as
which provides a good approximation for small .
If all nodes transmit with probability (including the desired one) and the receiver process has intensity , we have , and
where the factor is the probability that the transmitter under consideration is allowed to transmit times in a row.
5.2The local delay and the critical probability
Let the local delay be defined as
It denotes the time until the first successful transmission (starting at time ). For a deterministic link distance, we have
and the delay distribution is
The mean local delay or simply mean delay can be expressed as
While this sum cannot be directly evaluated, the mean can be obtained using the fact that outage events are conditionally independent given , i.e., by taking an expectation of the inverse conditional Laplace transform of the interference, see . This yields
So for a deterministic link distance, the mean delay is finite for all .
For random (but fixed) link distance, the mean delay is analogously expressed as
where can be expressed using the joint success probabilities from Cor. ?. It turns out that in this case, it is not guaranteed for to be finite for any . In fact, it was shown in  that if and only if
where as above.
Here we would like to explore whether this phase transition, i.e., the existence of a critical transmit probability such that for , is mainly due to the random link distance or due to the interference correlation. The following corollary establishes the condition for finite mean delay if interference was independent.
So even if the interference was independent from slot to slot, the static random transmission distance would cause the local delay to become infinite if the spatial contention or the transmit probability are too large. The critical transmit probability is shown in Fig. ? for the cases of independent and dependent interference and different ratios as a function of for . The parameter in and strongly depends on . The two critical probabilities divide the range of into three regimes: For , the mean delay is always finite. For , the mean delay is finite only if the interference is independent. For , the mean delay is always infinite.
It can be seen that for (), , which indicates that in this regime, the divergence of the mean local delay is mainly due to the random transmission distance.
5.3Alternative expression of the mean local delay and a binomial identity
As mentioned above in , the mean delay can also be expressed as a sum of . The joint success probability, averaged over the link distance, is given in Cor. ?. With independent interference, the diversity polynomial is replaced by , and applying inclusion-exclusion to yields
where . The mean delay follows as
This is identical to , which implies that
This identity may be of independent interest.
Using , the delay distribution with independent interference can be calculated as follows.
The bound is obtained from a bound on the ratio of gamma functions . It is asymptotically exact as . It reveals that is a necessary and sufficient condition for a finite mean, reproducing the result in via a different approach.
5.4Mean local delay calculation based on Taylor expansion
Here we use the linear approximation from to calculate the mean delay. With
where is the hypergeometric function. will be the estimated mean delay.
Expanding the hypergeometric function, we have
The negative term goes to zero since the sum is bounded by . Again applying the bound from  and noting that it is asymptotically exact as ,
So for , we obtain