Capacity approaching codes for photon counting receivers Marina Mondina, Fred Daneshgaranb, Inam Baria, Maria Teresa Delgadoa, a Dept. di Elettronica, Politecnico di Torino, C.so Duca degli Abruzzi, 24, 10129 Torino, Italy b ECE Department, California State University, Los Angeles, California, US ABSTRACT In [1] a lowcomplexity photoncounting receiver has been presented, which may be employed for weakenergy optical communications and which is typically modeled through its equivalent Binary Symmetric Channel (BSC) model. In this paper we consider the scheme described in [1], we model it as a time varying Binary InputMultiple Output (BIMO) channel and analyze its performance in presence of softmetric based capacity approaching iteratively decoded error correcting codes, and in particular using softmetric based Low Density Parity Check (LDPC) codes. To take full advantage of such detector, soft information is generated in the form of LogLikelihood Ratios (LLRs), achieving reduction in Bit Error Rate (BER) and Frame Error Rate (FER) with respect to classical BSC and Additive White Gaussian Noise (AWGN) channel models. Furthermore, we explore the limits of the achievable performance gains when using photon counting detectors as compared to the case when such detectors are not available. To this end, we find the classical capacity of the considered BIMO channel, clearly showing the potential gains that photon counting detectors can provide in the context of a realistic costeffective scheme from an implementation point of view. Furthermore, we show that from a channel modeling point of view, we can observe that the BIMO channel can be approximated with an AWGN channel for high values of mean photon count Nc, while the AWGN model offers an unreliable result with a low mean photon number Nc, (i.e. with low raw BER). This effect is more evident with lower coding rates. Keywords: Softmetric, channel capacity, forward error correction codes, LDPC codes, photon counting receiver, weakenergy optical communications
1. INTRODUCTION Most long distance optical communication schemes today employ relatively weak laser sources with small mean photon count at the receiver and nonphoton number discriminating detectors with acceptable dark count rates and detector deadtimes. A step forward in the evolution of such schemes would be to employ photon counting detectors at the receiver. This article focuses on such a paradigm, and in particular on a transmission scheme based on a lowcomplexity photoncounting receiver proposed in [1] and reported in Figure 1. Given that one could use photoncounting receivers for weakenergy optical communications, the question arises as to how such a detector may be employed to improve the system performance. This question may be answered by determining the capacity of the corresponding optical channel, and the achievable residual Bit Error Rate (BER) on such a channel. From a telecommunication point of view, the presence of a photon counting detector provides the possibility of generating at the receiver a softmetric (as opposed to a hard metric which essentially indicates the presence or absence of a signal). In what follows, the potential improvements that may be obtained in terms of classical capacity and residual BER using the photoncounting detectors based scheme proposed in [1] will be investigated. In particular, the paper is organized as follows: the considered system is described in section 2, where the corresponding channel model and LLR metric are also defined. The associated channel capacity is derived in section 3, while the achievable residual BER in presence of LDPC coding is shown in section 4. The conclusions are drawn in section 5.
[email protected] ,
[email protected],
[email protected] Phone: +39 011 090 4023 www.tlc.polito.it/dcc_team/
[email protected] Phone: +1 323 3434480 http://www.calstatela.edu/faculty/fdanesh/fdanesh.htm
2. THE CONSIDERED SYSTEM In the considered scheme [1], the information bit k=0,1 is encoded by applying the unitary transformation ( ) to the polarization degree of freedom of a coherent state  ⟩, which is assumed to be initially in the polarization state  ⟩  ⟩  ⟩ . This technique is based on the use of a Phase Beam Splitter (PBS) and two photon counters. The scheme allows √ one to map the discrete bit value “k” to an optical polarization quantum bit (or qubit), but at the detection stage can produce a discrete set of real numbers, which can also be expressed in the form of LogLikelihood Ratios (LLR) which can be used for soft information processing. A possible experimental setup is shown in Figure 1, where the polarization degree of freedom of a coherent state  is associated to the information bit “k” according to the following encoding rule: k → φk 0 → π/4 1 → 3π/4 showing a phase shift of / 2 of the qubit associated with k=0 relative to the qubit associated with k=1. The transformation U exp k is then applied, where 3 is the Pauli rotation matrix. This kind of transformation k
2
3
can be realized by means of a potassium dihydrogen phosphate (KDP) crystal driven by a high voltage generator and corresponds to change of the polarization from linear to elliptical.
Figure 1. The considered lowcomplexity photoncounting scheme.
At the detection stage a measurement of the phase shift of the qubit should be performed. This can be implemented as depicted in Figure 1 by using a HalfWave Plate (HWP) placed in front of a Polarizing Beam Splitter (PBS) with two photoncounters providing the number of photons in the reflected and transmitted beams, denoted as n0 and n1 respectively. Let n=n0+n1 denote the total number of detected photons. We assume this is also the total number of transmitted photons (in the hypothesis that no photon is lost), which is a Poisson distributed random variable with mean value En N c 2 . Notice that, due to the weak energy of the transmitted optical signal, the value of N c is typically small. From the knowledge of the photon counts n0 and n1, the actual value of the phase shift can be obtained by using the Bayesian estimator [2]
est pB (  n0 , n)d E  n0 , n, 0
where,
pB (  n0 , n)
p(k 0  ) n0 p(k 1  ) n1 p(0  ) n0 p(1  ) nn0 N N
is the probability density function of the received phase shift given the fact that n=n0+n1 photons have been received and n0 photons have been counted at the “k=0” output of the PBS. Ν is a normalization factor such that
p
B
(  n0 , n)d 1 .
0
This scenario generates the equivalent channel model with binary input (the random variable k), and multilevel output (the couple of random variables {n0,n1}) shown in Figure 2, which will be discussed in what follows. 2.1 Evaluation of the LogLikelihood Ratios In softdecoding algorithms, LogLikelihoodRatios are typically required, which, for the channel model shown in Figure 2 can be defined as: (
)
[
( (
{ {
}) ] })
( )
where, ( {
})
(
{
})
( )
is the probability that the transmitted bit was "k" given the measurement pair(n0,n1). Using Baye's Theorem, equation Errore. L'origine riferimento non è stata trovata. can be rewritten as: (
)
[
({ ({
} }
) ] )
( )
Since coherent states are being used, the number of photons n0 and n1 measured at the two detectors are uncorrelated and, ( ) in particular, are distributed according to a Poisson statistic. Given , the average number of photons of type “k” detected at the detector is given by the expression [1]: ( )
where
(
)
(1)
  is the average number of photons of the input coherent state, and ( 
)
(
(
))
( 
)
(
(
))
(2)
where, to make the analysis more general, it is assumed that during propagation, the qubit undergoes a phase diffusion process whose amplitude is characterized by the parameter (he feasibility of this scheme and its experimental demonstration have been thoroughly investigated in [3]). We have therefore: ({
} )
( )
(
( )
) (
)
(6)
( ) (7) where, is the Poisson probability distribution. Then, substituting equation (6) into equation (3) the following expression for the LLR can be easily obtained: (
)
(
)
( )
( 
)
(8)
where ( 
)
(
( ))
( 
)
( 
)
(
( ))
The system described up to this point can be modeled as a Discrete Memoryless Channel (DMC), and more precisely a Binary InputMultiple Output (BIMO) channel, with binary input k and outputs ( ) (where n is a Poisson distributed random variable) as shown in Figure 2. The capacity of this channel is evaluated in the next section.
Figure 2. BIMO channel model of the considered system.
3. CAPACITY EVALUATION The quantum channel in ddimensions is often modeled as a completely positive trace preserving map Ψ. The most common channel model is the depolarizing channel which depends on one parameter λ mapping a mixed state in Сd into:
where I is the d×d identity matrix. For a general quantum channel, let ε denote the ensemble of input states, and M the measurement or a Positive Operator Valued Measure (POVM) {Ej} at the channel output. The input state ensemble, channel and measurement together define a classical noisy channel with transition probabilities: (
)
Defining the probabilities over the input state, which we will denote as X, a natural definition of classical capacity of the quantum channel would be: ( )
(
)
where Y is the output state and I(X;Y) is the Shannon mutual information. The complication in defining capacity of the quantum channel in contrasts with the classical channel really arises in connection with purely quantum mechanical effects which have no analogue in the classical domain, i.e., entanglement. In particular, in general it is reasonable to assume (which is in fact shown to be true) that the capacity of parallel copies of a quantum channel with entangled inputs may be larger than the sum capacity of each channel treated separately. It turns out that for the most common channel model, namely the depolarizing channel, entanglement buys nothing. We note that the closest analogue of the binary communication scheme proposed here is Binary Phase Shift Keying (BPSK) using coherent states. It is well known that for such a scheme the Dolinar receiver achieves nearly optimal results with capacity [4]: (
(
√
))
where, H2(.) is the binary Entropy function. This capacity is close to the ultimate capacity obtained using an as yet unknown optimal receiver:
(
(
))
The Dolinar receiver requires a complicated feedback system for its implementation, hence there is significant difference in the level of the complexity with respect to a photon counting receiver. Our discussion thus far has been general and focused on quantum channels as tracepreserving maps. We have however a much more humble pursuit in this article that is modeling an experimental setup using realistic offtheshelf components, and specifically calculating the traditional Shannon capacity of the link viewed as a probabilistic transition mechanism that maps input bits into possibly multilevel outputs used for detection. In this sense, we model our channel as a Binary InputMultilevel Output (BIMO) DMC, and our goal is to contrast the capacity of a system employing photon counting detectors to that of the equivalent Binary Symmetric Channel (BSC) resulting from reducing the photon counts into presence or absence of signals (i.e., hard decoding). As noted earlier, the sufficient statistic for detection with photon counting detectors is the count difference of detector 1 and 0, i.e., (n1n0). The outputs of the two counters are independently distributed Poisson random variables, i.e. ( ) ( )
where
( )
(  ( 
( ) ( )
As a consequence, the difference ( (
) )
( )
) is Skellam distributed, and we have 
)
( )
(
( )
)
( )
(
( )
)
 (
√(
( )
( )
))
where, k=0 or 1, and Im(.) is the modified Bessel function of the first kind and order m. Note that m itself is an integer that can be positive or negative. Plugging known values of the parameters, we get: ( ) ( )
( )
(  ( 
) )
( 
) ( 
( )
Specializing to the case “0 is transmitted and is mapped to
P(n1 n0 m  0 ) e
Nc
similarly for the case “1 is transmitted and is mapped to
” we get:
2 e 2 e 2
P(n1 n0 m  1 ) e
m/2
2 e 2 Im N c 1 2
m/2
2 e 2 Im N c 1 2
2
2 e 2 e 2
”: 2
Nc
)
Let X be the random variable associated with the transmitted phase (or the transmitted bit “k”) and Y be the channel output (i.e. our sufficient statistic ( )). Then, the formulas above give us the channel transition probabilities for the considered DMC. Using the classic definition of mutual information:
(
)
and noting that the input is binary with ( manipulation we have:
( ) )
(  )
(
(

)
(

)
)
(
, and
(
)
(
)
, after some
) (
) )
(
where, (√
)⁄(√ and
Finally, the conditional entropy based on two parameters,
(  )
∑
∑(
(
)
) (
)
( )
can be written as:
√
 (
 (
)
)
√
)
( (

( (

))
))
While our BIMO DMC is neither symmetric nor weakly symmetric, it is not difficult to show that the maximizing input probability distribution is uniform. Hence, p=0.5 maximizes the mutual information leading to channel capacity. To compare the capacity of the link employing photon counting detector to that of a simple detector signaling the presence or absence of signal, we need to specify how such a detector behaves. It is logical to assume that crossover probability of the equivalent BSC (i.e. the raw BER, denoted in the following as QBER) associated with such a receiver can be obtained as: ∑ ( ∑
(

) 
)
(

)
(

)
(9)
) Notice that when ( (which for low values of Nc happens often), the detector chooses at random between k=0 and k=1. . Figure 3 depicts the capacity of our BIMO DMC and its comparison to the equivalent Binary Symmetric Channel (BSC) channel in case of hard decision decoding as a function of the mean photon count in the case the phase diffusion parameter is zero. Figure 4 depicts the capacity of our BIMO DMC and its comparison to the equivalent BSC in case of hard decision decoding as a function of the phase diffusion parameter for three different values of the mean photon count. It can be observed that the considered BIMO DMC channel offers a capacity improvement over the equivalent BSC. This improvement could lead to a BER improvement when comparing the two channels in presence of an error correction code. This is investigated in the next section.
4. BER PERFORMANCE In this section the performance obtainable with Forward Error Correction (FEC) codes applied to the scheme of Figure 1 (whose equivalent channel model is shown in Figure 2) is presented. A block of k information bits plus the corresponding r redundancy bits generate one codeword of a block FEC code with rate Rc k k r , where the codeword length is n= k + r. In order to assess the potential gains that can be obtained using the softmetric of equation (8), we have conducted simulation studies using FEC coding with Low Density Parity Check (LDPC) codes simulated with three different quantum channel models, all with the same equivalent uncoded raw BER observed on the channel, denoted as QBER and evaluated as in (9). Figure 5 depicts three sets of simulation results. Each pair of curves is associated with residual Bit Error Rate (BER) and Frame Error Rate (FER) values after decoding. LDPC codes with the following parameters have been considered: a) k=500, r = 500 and code rate Rc = 0.5, b) k=252, r = 156 and code rate Rc = 0.61, c) =750, r = 250 and code rate Rc = 0.75.
Figure 3. Classical capacity of BIMO DMC(solid curve) compared to the equivalent BSC with transition probability QBER, as a function of mean photon count Nc.
Figure 4. Classical capacity of BIMO DMC and equivalent BSC with transition probability QBER as a function of phase diffusion parameter for three different values of Nc (solid line: Nc=9, dashdot line: Nc=6, dash line: Nc=3).
The black pair (labeled “QBSC”) is for the reference system used for comparison. In this system, we suppose not to use the additional information derived from the knowledge of and , and we use as equivalent channel model a simple Binary Symmetric Channel (BSC) with binary input X, binary output Y and crossover probability QBER, (so that (  ) (  ) ), whose LLR values can be expressed as [6]:
( )
( [ (
 ) ]  ) {
a)
(
)
(
)
b)
c) Figure 5. BER and FER simulation results for a LDPC code with a) k=500, r = 500 and code rate Rc = 0.5, b) k=252, r = 156 and code rate Rc = 0.6 1and c) =750, r = 250 and code rate Rc = 0.75, obtained with different models of the quantum channel: BSC (QBSC curves), AWGN (QAWGN curves) and BIMO DMC (QBIMO curves).
The blue curves represent the performance obtainable over a fictitious Additive White Gaussian Noise (AWGN) channel model with a Signal to Noise Ratio (SNR) that would yield the considered raw bit error probability QBER if a binary antipodal scheme were used for data transmission (curves labeled as “QAWGN”). The “QBIMO” curves represent the core result and is associated with the use of a quantum channel modeled as a BIMO DMC as shown in Figure 2 with equivalent uncoded bit error probability QBER and LLR metrics generated via photon counting according to equation (8). Notice that for the “QBIMO” curves the QBER parameter is actually the crossover probability of the equivalent BSC as defined in (9) for the BIMO channel. As is evident from the results, there is significant reduction in BER and FER values when using, for the photon counting receiver, the appropriate BIMO channel model and the associated LLR metrics derived in (8) instead of the simpler BSC model (notice that the “QAWGN” curves have been only derived for reference, since with the small number of considered photons the AWGN channel model would not be appropriate). The FERBER performance show the improvement that can be achieved with the use of the BIMO channel model and the associated LLR metrics. For instance in Figure 5a ( Rc 0.5 ) at QBER=0.1, there is a more than three orders of magnitude improvement in BER and FER when comparing the proposed softmetric processing versus the reference protocol whereby the quantum channel is a BSC. Figure 5b and Figure 5c compare the FERBER performance of the considered channel models and LLR values with LDPC codes with rates 0.61 and 0.75, respectively. It can be observed that a FERBER improvement of up to several orders of magnitude can be obtained in these cases as well. It can also be observed that, at higher code rate, the performances obtainable with the BIMO LLR softmetrics gets closer to the performance of the classic AWGN channel metrics (although, as mentioned before, the AWGN model would not be applicable in the current case). Figure 7 compares the BER values obtained with the BSC and the BIMO channel models for different code rates, showing that, as expected, for higher rates a lower QBER values is required before significant coding gains can be observed.
Figure 6. BER simulation results for a LDPC code with code rate Rc = 0.5,0.61 and 0.75 obtained with different models of the quantum channel: BSC (QBSC curves), and BIMO DMC (QBIMO curves).
Figure 7. BER simulation results for a LDPC code with code rate Rc = 0.4 , 0.61 and 0.75 obtained with different models of the quantum channel: AWGN (QAWGN curves), and BIMO DMC (QBIMO curves).
From Figure 7, we can observe that the BIMO channel can be approximated with an AWGN channel for high values of Nc, i.e. low values of QBER, while the AWGN model offers an unreliable result with a low number of photons Nc, i.e. with high QBER. This effect is more evident with lower coding rates (as shown below, where the coding gain becomes apparent at low values of Nc).
Figure 8. BER simulation results for a LDPC code with code rate Rc = 0.5,0.61 and 0.75 obtained with BIMO channel for different values of Nc .
Finally, Figure 8 shows the residual BER on the BIMO channel BER for different values of Nc. CONCLUSIONS In this paper a lowcomplexity photoncounting receiver has been considered, which may be employed for weakenergy optical communications, and the performances of softmetric LDPC codes on the considered scheme have been presented. Different channel models for the considered system and the associated LLR metrics have been compared, proposing a time varying Binary InputMultiple Output (BIMO) channel and calculating the associated LLR metrics and channel capacity. It has been shown how the BIMO channel model outperforms a classic (and simpler) BSC scheme, taking full advantage of the additional information offered by the lowcost photon counting detector. We also showed
that the BIMO channel can be approximated with an AWGN channel for high values of Nc , i.e. low values of the raw bit error probability QBER, while the AWGN model offers an unreliable result with a low number of photons Nc, i.e. with high values of QBER.
REFERENCES [1] Olivares S. et alii, “Toward a ‘soft’ output quantum channel via Bayesian estimation,” Proc. ISABEL 2010, Rome (Italy), Nov. 710, (2010). [2] Teklu, B., Olivares S. and Paris, M. G. A., “Bayesian estimation of oneparameter qubit gates,” J. Phys. B: At. Mol. Opt. Phys. 42, (2009). [3] Brivio, D. et alii, “Experimental estimation of oneparameter qubit gates in the presence of phase diffusion,” Phys. Rev. A 81, (2010). [4] C.W. Lau, V. A. Vilnrotter, S. Dolinar, J. M. Geremia, and H. Mabuchi, “Binary Quantum Receiver Concept Demonstration,” IPN Progress Report 42165, May 15, (2006). [5] S. J. Dolinar, Jr., “An Optimum Receiver for the Binary Coherent State Quantum Channel,” MIT Research Laboratory of Electronics Quarterly Progress Report111, Institute of Technology, Cambridge, Massachusetts, pp. 115–120, October 15, (1973). [6] Mondin M., Delgado M., Mesiti F., Daneshgaran F., “Softprocessing for Information Reconciliation in QKD Applications,” International Journal of Quantum Information, vol. 9, pp. 155164, (2011).