A REDUCED COMPLEXITY TWO-DIMENSIONAL BCJR DETECTOR FOR HOLOGRAPHIC DATA STORAGE SYSTEMS WITH PIXEL MISALIGNMENT

Transcription

1 A REDUCED COMPLEXITY TWO-DIMENSIONAL BCJR DETECTOR FOR HOLOGRAPHIC DATA STORAGE SYSTEMS WITH PIXEL MISALIGNMENT 1 S. Iman Mossavat, 2 J.W.M.Bergmans 1 iman@nus.edu.sg 1 National University of Singapore, Singapore 2 Technische Universiteit Eindhoven, The Netherlands ABSTRACT As density in holographic data storage (HDS) systems increases, inter-symbol interference increases, and signalto-noise ratio (SNR) decreases. In order to combat these adverse effects, iterative reception techniques are of great interest because of their ability to achieve near-optimal bit-error-rate (BER) performance at low SNRs. Softdecision detectors are an integral part of such reception techniques. Existing soft-decision detectors for HDS are highly complex. In this paper, we extend an existing reduced-complexity BCJR detector so as to fit the characteristics of the HDS channel. The complexity reduction is achieved by exploiting the separability of the HDS channel. In order to further limit the complexity we design a novel partial response signal. An important added advantage of our technique is that it can handle high levels of pixel misalignment. 1. INTRODUCTION Holographic data storage (HDS) records/retrieves information in the form of two-dimensional holograms. Each hologram consists of millions of data bits. The entire information in a hologram is recorded / retrieved with a single flash of light which implies high data rates. Many holograms can be recorded throughout the volume of a thick material, allowing for high densities to be achieved. For a laser beam of wavelength λ, storage densities of order 1/λ 3 are predicted. This means that densities near to 1 TB/cm 3 are achievable for visible laser light. Holographic data storage poses challenging signal processing problems. Detection techniques should deal with the inherent non-linearity of this channel as well as with two-dimensional (2-D) inter-symbol interference (ISI). In HDS, data bits to be stored are presented by a spatial light modulator (SLM) during the recording phase. The SLM modulates the magnitude of the laser light. Later, a charge-coupled device (CCD) is used to detect the intensity of the laser light. This introduces a severe nonlinearity into the channel. In HDS, signal-to-noise ratio (SNR) is inversely proportional to the number of holograms stored throughout the volume of the recording media [1]. Consequently, low SNRs are inevitable as density increases. Accordingly, developing soft decision detectors for HDS channels is of fundamental importance, since such detectors are readily integrable with lowdensity parity-check (LDPC) codes into iterative reception schemes that can achieve near-optimal bit error rate (BER) performance at low SNRs. Most existing detection techniques for HDS [2], [3], [4] produce hard decisions and hence do not fit this bill. A notable exception is [5]. Unfortunately, complexity of the detector of [5] tends to be very high. In this paper we develop a much simpler yet near-optimum reception technique that is based on an extension of [6]. In [6] a reduced-complexity BCJR detector for a specific class of linear 2-D channels was described. BCJR is an optimal symbol-by-symbol maximum a posteriori detector that produces soft decisions. For an ISI span of L L the detector complexity in [5] is exponential in L 2 while complexity of the detector in [6] is exponential in L. The complexity reduction of [6] applies to linear channels that are separable, i.e. for which the 2-D ISI can be viewed as a concatenation of ISI along the rows and the ISI along the columns. Our current work is based on the observation that the HDS channel, while non-linear, has a similar property. (We will discuss this property further in Section 2 when we introduce the channel model.) This key observation permits us to extend the 2-D BCJR detector of [6] to deal with the non-linear nature of the HDS channel at no additional complexity. The resulting complexity is much lower than that of [5]. Even so, at high densities (i.e. for large ISI spans), it can still be very high. To limit complexity further, we resort to partial-response (PR) techniques that limit the ISI span prior to detection. To this end, we introduce a new target signal that involves the same non-linear mechanism as in the HDS channel. A further important challenge in HDS is pixel misalignment. If we assume an equal number of pixels on SLM and CCD, it is ideal for the pixels on these two devices to be spatially matched, i.e. each pixel on the CCD is exactly in front of the corresponding SLM pixel. In prac-

2 d y j j I i, j Row ISI (u) Column ISI (v). + 2 DCM: H = uv T Holographic Channel n i, j Figure 1: Channel model of HDS channel. tice it is impossible to achieve perfect pixel alignment due to a variety of adverse factors. The effect of pixel misalignment is substantial. As [1] showed, pixel misalignment significantly deteriorates the detection performance and even sometimes brings the achievable density to zero. Reference [1] presented a non-linear algorithm to mitigate the effects of pixel misalignment. However, this algorithm fails to work with an acceptable performance for high levels of pixel misalignment. We extend the well known discrete magnitude-squared channel model [3] for misaligned channels. The extended model has the same structure as the original model, with similar separability properties. Hence, we can still use our reception technique. Unlike the approach of [1], our technique works well even for severe misalignments. Furthermore, [1] uses decision feedback; consequently, error propagation may arise at low SNRs. Conversely, combinations of BCJR detectors and LDPC codes are known for their excellent performance at low SNRs. We also test linear PR targets along with the 2-D BCJR detector of [6], and observe a poor BER performance for severe misalignment. This clearly illustrates the necessity of accommodating non-linearity in the 2-D BCJR for HDS channels. The rest of the paper is organized as follows: In Section 2 we present the channel model and briefly discuss HDS channel properties. In Section 3 we discuss our reception technique and the modified separability properties. In Section 4 we present our PR signal and in Section 5 we discuss the corresponding equalizer and target optimization. In Section 6 we discuss the numerical results and finally we present our conclusions in Section 7. As there was no model for the case of HDS channels with pixel misalignment, we extended current models to deal with misalignment. We briefly state our results in the appendix. 2. CHANNEL MODEL We use the discrete magnitude-squared channel model of [3] to simulate CCD read-back values in HDS. We consider detector electronics noise only, which is zero-mean, additive, white, Gaussian (AWG). In other words we assume an electronics-noise dominated channel. We also extend the model of [3] to the case where pixel misalignment exists. The mathematical derivation of the extended channel model is presented in appendix A Figure 2: Discrete channel matrix for the pixel-aligned channel H 0.0,0.0 (Up) and the pixel-misaligned channel H 0.5,0.5 (Down). Figure 1 illustrates the model for the HDS channel. We denote the input data bits by d j. These bits pass through a linear 2-D ISI channel characterized by a discrete channel matrix (DCM) H. The ISI span is (2L+1) (2L+1) which means that (2L + 1) 2 pixels interfere for each readback value of the CCD. As shown in the appendix, H is separable, i.e. H = uv T (1) where u and v are (2L + 1) 1 vectors and v T is the transpose of v. Separability of the DCM H allows us to consider 2-D ISI as the concatenation of two channels representing row and column ISI respectively. First, data bits d j pass through the row ISI channel, characterized by u, and an intermediate output y j is produced. Then the symbols y j pass through the column ISI channel, characterized by v, and the magnitude of the result is squared to generate the noiseless channel output. White Gaussian noise n j is added afterward to produce the CCD read-back value I j. Note that we only observe I j, since y j is an intermediate signal. Since the DCM depends on the amount of misalignment in x and y directions, we actually use the notation H δx,δy = [h δx,δy j ] instead of H for representing the DCM, where δ x and δ y represent the pixel misalignment between the SLM and CCD. Now we can write the CCD read-back signal I j as follows: I j = h δx,δy j [d j ] 2 + n j (2) where denotes 2D convolution. From this point on, we will refer to a HDS by its corresponding DCM. We consider two HDS channels. The first channel is a perfectly pixel-aligned HDS channel, referred to as pixel-aligned HDS. Second channel is a HDS

3 channel with half pixel misalignment in both x and y directions, simply referred to as pixel-misaligned HDS. Figure 2 illustrates the DCM of these channels. The entry at the center of the DCM is presented in bold face. All the other entries correspond to interfering pixels. Note that H 0.5,0.5 represents the most severe case of pixel misalignment where we can see that 4 pixels contribute almost equally to the read-back value. 3. RECEPTION TECHNIQUE As shown in Figure 3, our reception technique is comprised of a linear minimum-mean-squared-error (MMSE) equalizer and the 2-D BCJR detector. Linear MMSE equalization for HDS channels was used before by [2]. First, we equalize the channel output, I j. The equalizer output s j is an estimate of the target signal s j to be described in the next section. For the time being, it is enough to know that the HDS channel and the target signal follow the same model and the only difference between them is the size of their DCM. The BCJR detectors we use here are based on the 2-D BCJR detectors in [6]. Detectors in [6] were developed for linear AWGN channels, and we modify them to work with the non-linear HDS channel. As Figure 1 suggests, the DCM is separable. First we use a BCJR detector similar to the column detector in [6] to produce log-likelihood ratio (LLR) values for the intermediate signal y j. We only have to modify the branch values of the detector trellis based on the non-linearity of the HDS channel. The remaining detector equations are kept unchanged. In Figure 3, we denote the output of the column detector by L yj. The column detector then passes L yj to another binary BCJR detector to compute the LLR of data bits, denoted by L dj. We refer to this detector as the row detector and its structure is exactly the same as the row detector in [6]. We decide on the bit values based on the sign of L dj. 4. NEW MAGNITUDE-SQUARED PARTIAL RESPONSE SIGNAL In order to limit the complexity of the 2D-BCJR we choose a partial response signal that has an S S support for S < 2L + 1. We present the following 2-D signal s j = γ j [d j ] 2 (3) as the equalization target. The structure of Equation 3 is identical to that of Equation 2, where γ j are target coefficients that control the shape of the target signal. We present the target coefficients γ j in matrix form, and we constrain the matrix to be separable, i.e. Γ = [ γ j ] = xy T. (4) I j Linear S % y j Column L i, j MMSE Detector Equalizer 2-D BCJR detector Row Detector, L d i j Figure 3: Reception for non-linear separable channel. Vectors x and y are S 1. For the rest of this paper, we refer to such a target signal by its underlying matrix Γ = [ γ j ]. Non-linearity is incorporated so that the signal can to be very close to the channel output. Hence, less equalization effort is needed and better noise-whitening is achieved. This will improve the performance of the BCJR detector. Furthermore, since Γ is separable, we can still use the simplified BCJR detector of Figure 3. We equalize the channel to a target signal with a support size of 2 2. Given this support size, the column detector traverses a non-binary trellis with four states and the row detector traverses a binary trellis with two states. 5. EQUALIZER AND TARGET OPTIMIZATION The error signal between the equalizer output and the corresponding target signal is e j = s j s j. (5) In order to derive the expression for mean-squared-error (MSE) and the optimal equalizer coefficients, we use a vector format to represent the variables. As [2] suggests, we represent equalizer, target, and their inputs and outputs by vectors instead of matrices. We can represent a matrix by a vector using any arbitrary convention. Assume that the equalizer support size is (2Q + 1) (2Q + 1). Vector c (2Q+1)2 1 represents the equalizer coefficients and I (2Q+1)2 1 is the equalizer input. Hence, the equalizer output (2D-BCJR input) s j is s j = c T I. (6) If we denote the target coefficients by Λ S 2 1, and the target input data bits by d S 2 1, we have: The MSE is s j = (Λ T d) 2 = Λ T dd T Λ. (7) ξ Λ = E [ (Λ T dd T Λ c T I) 2] = E [ (Λ T dd T Λ) 2] + c T Rc 2c T P Λ Λ (8) where R = E [ II T ] and P Λ = E [ IΛ T dd T ]. Although a non-linear target is used, ξ Λ is still convex in terms of c for a given Λ. So we take the gradient with respect to c and obtain c ξ Λ = 2Rc 2P Λ Λ. (9)

4 Setting this gradient to zero and solving for c we get c = R 1 P Λ Λ. (10) Note that P Λ depends on Λ. This dependency is not desirable if we wish to compute equalizer coefficients for various targets. However, there is a simple way to overcome this problem. Assume that vector Λ is expressed as a linear combination of some basis {v i } that contains S 2 linearly independent vectors, Λ = a i v i (11) S 2 i=1 where a i are scalars. Then P Λ = a i P vi (12) S 2 i=1 where P vi = E [ Iv ] i T ddt. Hence, if we compute P vi for the entire basis, we can efficiently compute the P Λ for any vector Λ of length S 2 1. For example, consider H 0.5,0.5 which is the DCM of the pixel-misaligned HDS in Figure 2. This matrix has four entries that are significantly larger than other entries. Consequently, the channel output is mostly dominated by 4 bits corresponding to these entries. So the 2 2 target coefficient matrix ( ) Γ CT = (13) which is simply obtained by truncating H 0.5,0.5 is intuitively a promising candidate. We refer to target signals with such a coefficient matrix as Channel Truncation (CT) target signals. As we have not yet developed an analytical way to find the optimal target, we perform a brute force search for a coefficients that yield the best BER performance. We search the space of 2 2 separable matrices; such matrices have the general form ( ) a 2 ab ab b 2 (14) where a and b are scalars. In order to limit the search complexity, we constrain a to be 1 and b to be smaller than 1. We increase b from zero to one and estimate the corresponding BER by simulation. This will largely reduce the search complexity, but it may lead to loss of optimality as well. In spite of posing these constraints, our results in the next section still illustrate that magnitude-squared targets achieve superior performance. A separate b is chosen for each SNR. Here, SNR is defined as ( ) 1 SNR = 10 log (15) σ 2 n BER BER - Pixel-aligned HDS Non Linear Target (Search) Non Linear Target (CT) Linear Target (Search) MMSE + Threshold SNR (db) Figure 4: BER performance of BCJR detection with linear and non-linear PR targets, and MMSE-threshold detection for pixel-aligned HDS. BER BER - Pixel-misaligned HDS Non Linear Target (Search) Non Linear Target (CT) Linear Target (Search) MMSE + Threshold SNR (db) Figure 5: BER performance of BCJR detection with linear and non-linear PR targets, and MMSE-threshold detection for pixel-misaligned HDS. where σ 2 n is the electronics noise variance. It is worthwhile to note that for electronics-noise dominated channels, σ n is proportional to number of recorded pages as [1] stated. 6. NUMERICAL RESULTS In our simulations, we use unity fill factors for SLM and CCD, normalized pixel width, Nyquist aperture width, and SLM contrast ratio of 100. A MMSE equalizer of kernel size 5 5 is used for all equalizations. We present the BER performance in Figures 4 and 5. We have also plotted the BER performance of BCJR with a linear 2- D PR target and the BER performance of a full response equalizer with threshold detection. For convenience we refer to the best target found as op-

5 timal target. We should bear in mind that because of constraining the search space, our results are not optimal, still they show the significant gains of using magnitudesquared target signals. Also note that the discrete channel matrices of the two HDS channels we study are diagonally symmetrical as Figure 2 suggests. We can see that the optimal non-linear target gives the best performance among different reception techniques/targets. For the pixel-aligned channel, the CT target is far away from optimality at low SNR. However, the performance gap between the optimal non-linear target and the CT target reduces at high SNR for pixel-aligned channel. For the pixel-aligned channel the CT target outperforms the optimal linear target at high SNR. For the pixel-misaligned channel the CT target always offers superior BER performance. In fact, for the pixel-misaligned channel, threshold detection and linear target PR fail due to the high amount of ISI. However, for non-linear targets the BER decays slowly and reaches a floor beyond SNR of 36 db. 7. CONCLUSION We extended the low-complexity, 2-D BCJR detector of [6] to the non-linear HDS channels. We exploited the separability property of the holographic data storage channel for this purpose. With simple adjustments, our 2-D BCJR detector is able to handle channel non-linearity at no additional complexity. We present a new partial response target signal that mimics the non-linear behavior of the channel. This new partial response enables us to detect at low complexity even in the face of severe pixel misalignment. By comparison, linear targets fail when severe misalignment exists. Appendix A. DERIVATION OF MISALIGNED CHANNEL MODEL Here, we present a concise mathematical derivation of the system model using similar notations and assumptions as [3]. We take page-wide pixel misalignment into account as we derive the model. Let us denote the binary data by d j and the SLM finite contrast ratio by ɛ. SLM represents one and zero binary values by two amplitude levels of 1 and 1/ɛ respectively. We choose ɛ = 100 in our simulations. We assume that SLM pixels have a rectangular shape with fill factor equal to one. We denote the pixel width of the SLM by s. After passing through the channel, the optical wave-front at the CCD is: z(x, y) = k,l d k,l h(x k s, y l s ) (16) where h(x, y) captures the physical characteristics of the HDS channel. In detail: h(x, y) = (D/λf L ) 2 s/2 s/2 s/2 s/2... sinc( (x τ 1)D λf L )sinc( (y τ 2)D λf L )dτ 1 dτ 2. (17) We assume a square aperture and denote its width by D. We also denote the laser light wavelength by λ, and the lens focal length by f L. In our simulations, we always assume Nyquist aperture width: D = D N = λf L s. (18) Note that h(x, y) is separable in terms of x and y, i.e. where h(x, y)=h 0 (x)h 0 (y), (19) h 0 (x)= D s/2 sinc( (x τ 1)D )dτ 1. (20) λf L s/2 λf L The CCD array integrates the incident intensity z(x, y) spatially and temporally to produce the read-back signal I j. In order to model the misalignment of the CCD and the SLM, we assume that the the CCD integrates a shifted signal z(x + δ x, y + δ y ). We restrict our study to global misalignments; so δ x and δ y represent page-wide misalignments in the x and y directions respectively. Furthermore, we denote the CCD pixel width by c. Consequently, the output intensity sequence I j is as follows: I j = i c+ c/2 j c+ c/2 i c c/2 j c c/2 z(x + δ x, y + δ y ) 2 dydx (21) where we assume that the CCD pixels have a rectangular shape with fill factor equal to one. We assume s = c =. (22) Now we show how to compute Equation 21 efficiently using same techniques as [3]. Plugging Equations 16 and 22 into Equation 21 we get: + /2 I j = + /2 d k,l d m,n... k,l m,n /2 /2 h(x + (i k) + δ x, y + (j l) + δ y ) h(x + (i m) + δ x, y + (j n) + δ y )dydx. (23)

6 Since h(x, y) is separable in terms of x and y, we can simplify Equation 23 further: I j = k,l m,n where G δ k,m is defined as: G δ k,m = + /2 /2 d k,l d m,n G δx i k,i m Gδy j l,j n (24) h 0 (τ +δ+k )h 0 (τ +δ+m )dτ. (25) This is very similar to [3] with one difference: now G δ = [G δ k,m ] depends on the misalignment value δ. These equations imply that for each value of δ along either axis, there is a corresponding matrix G δ. We proceed to simplify Equation 24 further. Using eigenvalue decomposition techniques, we can write G δ as G δ = i λ i q i q T i (26) where q i is the i th eigen vector of G δ and λ i is its corresponding eigenvalue. Note that q i is a column vector. We approximate G δ as: G δ λ δ maxq δ λ max ( q δ λmax ) T (27) where λ δ max is the largest eigenvalue and q δ λ max is its corresponding eigenvector. From now on, we denote these quantities by λ δ and q δ respectively. We denote the i th entry of q δ by ( q δ). Note that i G δx k,m Gδy l,n λ ( λδx δy q δx) ( k q δ x ) ( m q δ y ) ( l q δ y ). (28) n Now let us define m,n = ( λ δx λ δy q δ x (q )m δy ) n (29) H δx,δy which yields: G δx k,m Gδy l,n Hδx,δy k,l Equation 29 implies that H δx,δy is separable: H δx,δy m,n. (30) where denotes 2-D convolution. Basically, Equation 33 implies that in the presence of page-wide misalignments, the structure of the channel model is unaltered. It is worthwhile to note that by storing N different vectors of q δi we can construct channel models for N 2 combinations of misalignments along x and y directions. In order to show that Equation 33 is an accurate approximation, we computed the I j values based on Equation 24 and Equation 33 respectively. The normalized mean squared error is at most 7%. 8. REFERENCES [1] L. Menetrier and G.W. Burr, Density implications of shift compensation postprocessing in holographic storage systems, Applied Optics, vol. 42, no. 5, pp , [2] K.M. Chugg, X. Chen, and M.A. Neifeld, Twodimensional equalization in coherent and incoherent page-oriented optical memory, Journal of the Optical Society of America A, vol. 16, no. 3, pp , [3] M. Keskinoz and B.V.K.V. Kumar, Discrete Magnitude-Squared Channel Modeling, Equalization, and Detection for Volume Holographic Storage Channels, Applied Optics, vol. 43, no. 6, pp , [4] A. He and G. Mathew, Nonlinear equalization for holographic data storage systems, Applied Optics, vol. 45, no. 12, pp , [5] X. Chen, KM Chugg, and MA Neifeld, Near- Optimal Parallel Distributed Data Detection for Page- Oriented Optical Memories, IEEE Journal of Selected Topics in Quantum Electronics, vol. 4, no. 5, pp , [6] Y. Wu, J.A. OSullivan, N. Singla, and R.S. Indeck, Iterative Detection and Decoding for Separable Two- Dimensional Intersymbol Interference, IEEE Transactions on Magnetics, vol. 39, no. 4, pp. 2115, H δx,δy = λ δx λ δy q δx q δy T (31) Finally, we simplify Equation 24 as follows: I j k,l m,n d k,l d m,n H δx,δy k,l Hm,n δx,δy (32) = H δx,δy [d j ] 2. (33)