2D Sparse Sampling Algorithm for N D Fredholm Equations with Applications to NMR Relaxometry Ariel Hafftka Hasan Celik Applied Mathematics and Scientific Computation University of Maryland, College Park ahafftka@math.umd.edu Laboratory of Clinical Investigation, National Institute on Aging National Institutes of Health hasan.celik@nih.gov Alexander Cloninger Wojciech Czaja Richard G. Spencer Applied Mathematics Yale University alexander.cloninger@yale.edu Department of Mathematics University of Maryland, College Park wojtek@math.umd.edu Laboratory of Clinical Investigation, National Institute on Aging National Institutes of Health spencerri@mail.nih.gov Abstract—In [1], Cloninger, Czaja, Bai, and Basser developed an algorithm for compressive sampling based data acquisition for the solution of 2D Fredholm equations. We extend the algorithm to N dimensional data, by randomly sampling in 2 dimensions and fully sampling in the remaining N − 2 dimensions. This new algorithm has direct applications to 3-dimensional nuclear magnetic resonance relaxometry and related experiments, such as T1 −D−T2 or T1 −T1,ρ −T2 . In these experiments, the first two parameters are time-consuming to acquire, so sparse sampling in the first two parameters can provide significant experimental time savings, while compressive sampling is unnecessary in the third parameter. I. I NTRODUCTION We consider the problem of solving discrete separable Fredholm integral equations of the form m = (K1 ⊗ · · · ⊗ KN )f + e, (1) where m is the observed data vector of length m1 × · · · × mN , f ≥ 0 is a nonnegative distribution vector of length n1 × · · · × nN to be solved for, e is an unknown small error vector of length m1 · · · mN , and each Ki is a known mi × ni matrix. The problem (1) can be rewritten in terms of N D arrays (tensors) in the form M = (K1 , . . . , KN ) · F + E (2) where M and E are of size m1 × · · · × mN , F is of size n1 × · · · × nN and (K1 , . . . , KN ) · F denotes the result of multiplying F by Ki along the ith axis, for i = 1, . . . , N . The tensors M , E, and F are obtained by arranging the entries entires of m, e, and f into tensors lexicographically. In nuclear magnetic resonance relaxometry and related experiments, M denotes the experimentally acquired data and F denotes the unknown distribution of specific parameters characterizing a sample. Parameters of interest include, for example, spin lattice relaxation time (T1 ), spin-spin relaxation time (T2 ), and diffusion coefficient (D). T1 and T2 indicate the rate at which perturbed magnetization returns to equilibrium in the longitudinal and transverse planes, providing information about molecular composition and mobility and microscopic structure, while D indicates translational mobility. In conventional 1D NMR experiments, for example a T2 experiment, F is a column vector giving the distribution of c 978-1-4673-7353-1/15/$31.00 2015 IEEE T2 values in the sample. A given sample can be characterized by its distribution of relaxation times [9]. Higher dimensional NMR experiments aim to compute the joint density function F of one or more parameters. For example, in a T1 - T2 experiment, F denotes the joint distribution of T1 and T2 . These 2D experiments have seen growing applications in the chemical and biological sciences and permit a much more complete description of materials [10] [11]. Given the success of 2D relaxometry and related experiments, it is clear that the ability to determine the joint density function of multiple parameters in a sample provides tremendous analytic power. Higher dimensional experiments have also been observed to exhibit improved recovery stability [4]. It would therefore be of great value to have available higher dimensional NMR experiments for materials and tissue characterization. However, each additional dimension results in a substantial increase in data acquisition time. Compressive sensing (CS) is a mathematical theory based on the idea that if a signal is sparse in some basis, it can often be accurately recovered from a small set of incoherent measurements [5] [6]. For the case N = 2, Cloninger, et. al, developed CS based algorithm for the solution of (2) from observations of M on a random subset of its entries [1]. While there have been extensive applications of CS to magnetic resonance imaging (MRI) [8] using various types of sparsity, such as with respect to L1 , TV, and wavelet bases, we are not aware of any previous applications of CS to NMR relaxometry or related experiments other than Algorithm 1 in [1]. Unlike MRI, which requires Fourier methods, relaxometry problems require the solution of an inverse Laplace transform. One natural way to extend the algorithm in [1] to N D would be to randomly sample M . Filling in the missing entries of M is a low-rank tensor completion problem, for which several algorithms have recently been developed [12] [7]. However, for some experiments, such as T1 - D - T2 , random sampling of M along the axis corresponding to T2 does not provide significant experimental time savings, so randomly sampling of M is not an efficient sampling strategy. If we randomly sample M in two axes and fully sample in the remaining axes, the problem naturally splits into independent matrix completion problems. With this motivation, we develop an extension of the algorithm in [1] that uses random sampling along the first 2 axes of M and full sampling along the remaining N − 2 axes of M . Section II establishes notation, Section III summarizes a standard algorithm for the solution of relaxometry and related problems, and Section IV restates the 2D reconstruction algorithm from [1]. Section V generalizes the 2D algorithm to N Dimensions using sampling along 2D slices. In Sections VI and VII we state the error bound from [1] and prove a similar bound for the N D algorithm. Sections VIII and IX describe results on simulated and experimental data. II. N OTATION An N -tensor M of size m1 × · · · × mN is an element of Rm1 ×··· ,mN , i.e., an N D matrix, with entries denoted by M [i1 , . . . , iN ]. We lexicographically order the indices of an N -tensor by the condition that (i1 , . . . , iN ) < (j1 , . . . , jN ) ⇐⇒ i1 < j1 or for some 1 ≤ k < N, i1 = j1 , i2 = j2 , . . . , ik = jk and ik+1 < jk+1 . We define rowvec(M ) to be the length m1 m2 · · · mN column vector obtained by ordering the entries of M lexicographically. For example, if M is a matrix, rowvec(M ) is the column vector obtained by concatenating the rows of M . Given a column vector v of length m1 · · · mN , we define rowreshape(v, m1 , . . . , mN ) to be the N -tensor of size m1 × · · · × mN obtained by rearranging the entries of v into an N -tensor according to the lexicographical ordering. Hence rowreshape(rowvec(M ), m1 , . . . , mN ) = M. The Frobenius norm P of M is defined by ||M ||F = || rowvec(M )||2 = i1 ,...,iN |M [i1 , . . . , iN ]|2 . If M is a matrix, the operator norm ||M ||2 = σ1 (M ) is the large singular Prank(M ) value of M . The nuclear norm ||M ||? = i=1 σi (M ) is the sum of the singular values of M . If A is m1 × n1 and B is m2 × n2 , we define the tensor product of A and B, denoted A ⊗ B, to be the (m1 m2 ) × (n1 n2 ) matrix A[1, 1]B · · · A[1, n1 ]B .. .. .. . . . A ⊗ B := A[m1 , 1]B · · · A[m1 , n1 ]B For any fixed indices i3 , . . . , iN , we let M [·, ·, i3 , . . . , iN ] denote the matrix obtained by fixing the last N − 2 indices of M . Any vector obtained by fixing all indices except the k-th, i.e., of the form A[j1 , . . . , jk−1 , ·, jk+1 , . . . , jN ], is called a k-column of A. If Ω ⊂ {1, . . . , m1 } × · · · × {1, . . . , mN } is an arbitrary subset of the indices of M , we let M [Ω] be the vector obtained by listing the entries M [i1 , . . . , iN ] for which (i1 , . . . , iN ) ∈ Ω in the lexicographical ordering. The tensor product of N matrices Ki of size mi × ni , for i = 1, . . . , N , is a matrix of size (m1 · · · mN ) × (n1 · · · nN ) defined by iterating the previous definition. If F is an N -tensor of size n1 × · · · × nN and Ki are matrices of size mi × ni for i = 1, . . . , N , we define (K1 . . . , KN ) · F := rowreshape (K1 ⊗ · · · ⊗ KN ) rowvec(F ), m1 , . . . , mN . It can be shown that (K1 , . . . , KN ) · F is the N -tensor of size m1 × · · · × mN obtained by multiplying all 1-columns of F by K1 , all 2-columns of F by K2 , . . . , and all N -columns of F by KN , in any order. For example, if F is a matrix, (K1 , K2 ) · F = K1 F K20 is the result of multiplying all the columns of F by K1 and all the resulting rows by K2 . If Ji and Ki are mi × ni and ni × ri matrices for i = 1, . . . , N , (J1 ⊗ · · · ⊗ JN )(K1 ⊗ · · · ⊗ KN ) = (J1 K1 ) ⊗ · · · ⊗ (JN KN ). If Ki are matrices of size mi × ni for i = 1, . . . , N with singular value decompositions (SVDs) given by Ki = Ui Si Vi0 , the tensor product has SVD, it is shown in [22] that K1 ⊗· · ·⊗KN = (U1 ⊗· · · UN )(S1 ⊗· · ·⊗SN )(V1 ⊗· · ·⊗VN )0 . Let Ki = Ui Si Vi0 be the reduced SVD of each kernel. III. VSH A LGORITHM In applications, the kernels Ki have rapidly decaying singular values, so (2) is ill-conditioned. To reduce the sensitivity to noise, we solve for α > 0, the Tikhonov regularization min ||M − (K1 , . . . , KN ) · F ||2F + α||F ||2F . F ≥0 (3) Our N D algorithm and the 2D algorithm in [1] approximately solve (3) using the Venkatarmann, Song, Hurlimann (VSH) Algorithm developed in [2], which we summarize as follows. Since the singular values of each kernel decay rapidly, we assume that each Ki has low rank ri ≤ min(mi , ni ). This assumption is equivalent to truncating the singular values of each kernel, and thus improves the condition number of the kernel K1 ⊗ · · · ⊗ KN . Under this low-rank assumption, the problem (3) is equivalent to min ||M̃ − (K̃1 , . . . , K̃N ) · F ||2F + α||F ||2F , F ≥0 (4) 0 ) · M is of size where the compressed data M̃ := (U10 , . . . , UN r1 × · · · × rN and for i = 1, . . . , N , the compressed kernels K̃i := Ui0 Ki = Si Vi0 . are of size ri × ni . The VSH can be used to rapidly solve (4). Since the solution F of (3) and (4) only depends on the compressed data M̃ , Cloninger et. al. observed that any CS recovery approach should aim to recover M̃ , not the full data set M [1]. Once M̃ is recovered using compressive sensing, F can be obtained from the VSH Algorithm. IV. 2D R ECONSTRUCTION A LGORITHM The 2D algorithm in [1] aims to recover M̃ from measurements of M on a random subset of its indices. Cloninger, et. al., observed that since Ui0 Ui = Iri , M = (K1 , K2 ) · F = (U1 S1 V10 , U2 S2 V20 )·F = (U1 , U2 )·M̃ . Hence, for each (i1 , i2 ), we have M [i1 , i2 ] = (U1 [i, ·], U2 [j, ·]) · M̃ =< U1 [i, ·]0 U2 [j, ·], M̃ > . (5) As we shall see later, for our applications the rank 1 matrices {Ui [i, ·]0 U2 [j, ·]}i,j are highly incoherent, and thus provide a robust set of measurements from which to recover M̃ . There is extensive previous work showing that low rank matrices can often be accurately recovered by nuclear norm minimization, under certain favorable conditions, such as incoherence [14] [15] [16] [17] [13] [18] [19] [20]. We now state the algorithm in [1]. Let J = {1, . . . , m1 } × {1, . . . , mn }. Fix the number of measurements P ≤ |J|. Algorithm 1. (from [1]) 2D Reconstruction 1) Choose a sampling set Ω ⊂ J with |Ω| = P uniformly at random. 2) Let y[i1 , i2 ] = M [i1 , i2 ] + e[i1 , i2 ] denote noisy measurements of the entries of M , with ||e[Ω]||F ≤ . Only the entries y[i1 , i2 ] with (i1 , i2 ) ∈ Ω will be used. 3) Reconstruct M̃ by approximately solving the nuclear norm minimization min ||M̃ ||? . (6) || (U1 ,U2 )·M̃ −y [Ω]||2 ≤ 4) Solve (4) using the VSH Algorithm, starting with the data M̃ recovered in the previous step. An approximation solution of (6) can be rapidly obtained by fixed point continuation (FPC), a singular value thresholding algorithm [3]. V. E XTENSION TO 2 + N − 2 D IMENSIONS Let N ≥ 3. We will show that Algorithm 1 can be extended to N D by randomly sampling in the first 2 axes and fully sampling in the remaining N −2 axes. Let J = {1, . . . , m1 }× {1, . . . , m2 }, K = {1, . . . , m3 } × · · · × {1, . . . , mN }, and K̃ = {1, . . . , r3 } × · · · × {1, . . . , rN }. Let Ω ⊂ J with |Ω| = P as in Algorithm 1. We will show how M̃ can be recovered from observations of M [Ω × K]. For X ∈ Rm1 ×···×mN , define P(X) = 0 (Im1 , Im2 , U30 , . . . , Um ) · X, the compression of X along N axes 3, . . . , N . Observe that P(M ) = (U1 , U2 , Ir3 , . . . , IrN ) · M̃ . Hence, for each (i1 , i2 ) ∈ Ω and (j3 , . . . , jN ) ∈ K̃, P(M ) [i1 , i2 , j3 , . . . , jN ] (7) = (U1 [i1 , ·], U2 [i2 , ·]) · M̃ [·, ·, j3 , . . . , jN ]. These measurements of M̃ [·, ·, j3 , . . . , jN ] are of the same form as (5). Furthermore, P(M )[Ω×K̃] is directly computable from M [Ω × K]. Algorithm 2. N Dimensional Reconstruction 1) Choose Ω ⊂ J with |Ω| = P uniformly at random. 2) Let y[i1 , . . . , iN ] = M [i1 , . . . , iN ] + e[i1 , . . . , iN ] be measurements of M , with ||e[Ω × K]||F ≤ . 3) For each (j3 , . . . , jN ) ∈ K̃, reconstruct M [·, ·, j3˜, . . . , jN ] by approximately solving the nuclear norm minimization min ||M̃ [·, ·, j3 , . . . , jN ]||? . || (U1 ,U2 )·M̃ [·,·,j3 ,...,jN ]−P(y) [Ω]||2 ≤ 4) Solve (4) using the VSH Algorithm. VI. 2D R ECONSTRUCTION E RROR E STIMATE We summarize the error bound for (6) in the case N = 2, derived and proved [1], which rely on results on restricted isometry property (RIP) [20] for random incoherent measurements and bounds on nuclear norm minimization recovery error [14] under RIP. Definition 1. Let V be a Hilbert space. A collection {vi } ⊂ V is a bounded norm Parseval tight P frame for V if it is bounded and if for all x ∈ V , ||x||2 = i | < x, vi > |2 . The set of matrices V = Rr1 ×r2 is a Hilbert space with the Euclidean inner product. Since the columns of U1 and of U2 are orthogonal, it can be easily shown that {U1 [i, ·]0 U2 [j, ·]}(i,j)∈J is a bounded norm Parseval tight frame for V . Definition 2. (from [1], based on [20]) Let V = Rr1 ×r2 . A bounded norm Parseval tight frame {vj }j∈J for V with finite index set J is said to have incoherence parameter µ if for all 1 ,r2 ) . j ∈ J, ||vj ||22 ≤ µ max(r |J| A small incoherence parameter µ ensures that if X is a low rank matrix, the inner product magnitudes | < vj , X > | are not too concentrated on any small subset of indices [20]. Intuitively, this means that randomly chosen measurements < vj , X > are more likely to capture enough information to reconstruct X. For a Parseval tight frame in which each vj is rank 1, we have ||vj ||2 = ||vj ||F , and it follows that the incoherence parameter can be bounded above and below: |J| min(r1 , r2 ). (8) r1 r2 The maximum bound is attained by any frame containing a unit-vector while the lower bound is obtained, for example, by the Fourier frame F(i,j) [k, l] = √m11 m2 exp(2πik/m1 + 2πjl/m2 ). We will see later that in our applications, {U1 [i, ·]0 U2 [j, ·]}(i,j)∈J is highly incoherent. We now state the error estimate from [1]. For any matrix X, let Xr denote the best rank r approximation of X in the Frobenius norm. min(r1 , r2 ) ≤ µ ≤ Theorem 1. (from [1]) Assume that the Parseval tight frame {U1 [i, ·]0 U2 [j, ·]} has incoherence µ. Let r ≥ 0 and 0 < δ < 1/10. If the number of measurements P satisfies Cµ(5r) max(r1 , r2 ) log5 max(r1 , r2 ) log(P ) P ≥ , (9) δ2 then with probability greater than 1−exp(−C) over the choice of Ω, the solution M̃est obtained in Algorithm 1 satisfies P −1/2 ||M̃ − M̃r ||? √ ||M̃est − M̃ ||F ≤ C0 + C1 |J| r for all matrices M̃ , where C0 and C1 are small constants. VII. E RROR E STIMATE FOR 2 + (N − 2) D IMENSIONAL S AMPLING We now state our main result, which bounds the recovery error for Algorithm 2. If X is a tensor, we define Xr to be the tensor for which each Xr [·, ·, j3 , . . . , jN ] is the best rank r approximation of X[·, ·, j3 , . . . , jN ]. Theorem 2. Let N ≥ 3. With notation as in Algorithm 2, assume that the Parseval tight frame {U1 [i, ·]0 U2 [j, ·]} has incoherence µ. Let r ≥ 0 and 0 < δ < 1/10. If the number of measurements P satisfies (9), then with probability greater than 1 − exp(−C) over the choice of Ω, the solution M̃est obtained in Algorithm 2 satisfies The regularization parameter α = 1−10 . The singular values for the kernels were truncated so that the condition number was at most 104 . For each sampling percentage, Algorithm 2 was performed for 50 random choices of Ω. Each inversion F is considered admissible if there are three peaks, excluding small edge artifacts. The relative errors for M̃est and the peak centers of mass and integrals for the recovered distribution Fest are reported in Table I, with all relative errors computed with respect to the true data, averaged over admissible inversions. Fig. 1. Simulated 3-peak distribution F recovered with full sampling (left) and 5% compressive sampling on 2D slices (right). 100% sampling 5% sampling ||M̃est − M̃ ||2F ! 2 2C02 X ≤ M̃ − M̃r [·, ·, j3 , . . . , jN ] r ? + 2C12 N P −1 Y ri 2 |J| i=3 for all tensors M̃ , where the summation is over all (j3 , . . . , jN ) ∈ K̃ and C0 and C1 are as in Theorem 1. Proof. For each (j3 , . . . , jN ) ∈ K̃, the sampling operation in (7) is of the same form as in Algorithm 1. Hence we apply the result of Theorem 1 to P(y)[·, ·, j3 , . . . , jN ] in place of y and and M̃ [·, ·, j3 , . . . , jN ] in place of M̃ to obtain ||(M̃est − M̃ )[·, ·, j3 , . . . , jN ]||2F P −1/2 2 M̃ − M̃r [·, ·, j3 , . . . , jN ] ? √ + C1 ≤ C0 |J| r 2 P −1 2C02 M̃ − M̃r [·, ·, j3 , . . . , jN ] ? ≤ + 2C12 2 . r |J| Summing QN over (j3 , . . . , jN ) ∈ K̃ gives the conclusion, since |K̃| = i=3 ri . TABLE I 3- DIMENSIONAL SIMULATION 100% 25% 10 % 5% 2.5% 1% 0.0085 0.0027 0.0068 0.0086 0.0028 0.0070 0.0312 0.0158 0.0132 0.3934 0.3536 0.1257 0.0099 0.0131 0.0223 0.0083 0.0145 0.0221 0.0088 0.0149 0.0230 0.0233 0.0301 0.0414 0.2267 0.2211 0.3154 1 1 1 0.98 0.6 3.3e-5 5.8e-5 8.7e-5 0.0024 0.0663 Relative error in peak centroid Peak 1 Peak 2 Peak 3 0.0081 0.0024 0.0067 0.0081 0.0025 0.0069 Relative error in peak integral Peak 1 Peak 2 Peak 3 0.0092 0.0132 0.0217 Admissibility 1 Relative error in M̃ 0 VIII. 3-D IMENSIONAL S IMULATIONS IX. 3- DIMENSIONAL T1 -D-T2 OLIVE OIL EXPERIMENT We performed simulations on a 3-dimensional distribution F of size 64×64×64 with 3 hemispherical peaks of radius r = 0.1 and with centers c1 = (0.7, 0.3, 0.7), c2 = (0.3, 0.7, 0.3), and c3 = (0.3, 0.5, 0.7). The kernels are defined for i = 1, 2, 3 by Ki [j, k] = exp(−τi [j]/ti [k]), . We tested Algorithm 2 on a 3D T1 - D - T2 experiment obtained from an olive oil sample. For each sampling percentage, we report in Table II the relative error in M̃est and in the recovered peak centers of mass and integrals, averaged over the admissible results from 50 random sampling sets Ω. The data M is of size 64 × 64 × 128 and F is of size 64 × 64 × 64. The compressed data M̃ was of size 8 × 4 × 12. The inversion parameters were α = 10−9 and the singular values were truncated so that the resulting kernel has condition number at most 105 . For i = 1, 2, τi consists of mi = 128 logarithmically spaced points from 0.05 to 4, while τ3 consists of m3 = 1024 linearly spaced points from 0.05 to 4. For i = 1, 2, 3, ti consists of mi = 64 points linearly spaced from 0.05 to 1. The data M̃ has a signal to noise ratio (SNR) of approximately 256, where SNR = ||M̃ ||F /||Ẽ||F , and E is pseudorandom gaussian noise. Experimental details: Experimental data was collected on the olive oil sample at 25◦ C using a 400 MHz Bruker Avance III NMR spectrometer equipped with a 5 mm Micro2.5 micro-imaging solenoidal coil. The pulse sequence consisted of an inversion recovery module with variable inversion times, followed by a stimulated echo diffusion encoding with variable diffusion-sensitizing gradient strengths and a CPMG sequence with acquisition at echo maxima. Experimental parameters included: echo time TE = 2ms, number of echoes NE = 512, repetition time TR = 6s, number of inversions NI = 64 with inversion times sampled logarithmically between 50 and 3250ms, and 64 diffusion sensitization b-values logarithmically spaced between 1.25 and 5085s/mm2 , with a diffusion encoding period of ∆ = 20ms and bipolar encoding gradient duration δ = 1ms for each gradient value. Fig. 2. Experimental 2-peak T1 -D-T2 distribution F recovered with full sampling (left) and 5% compressive sampling on 2D slices (right). *Denotes inversion artifacts. 100% sampling 5% sampling 2 guarantees good recovery with high probability at very small sampling percentages, provided that µ grows slowly compared to |J|. XI. C ONCLUSION The 2D sparse sampling algorithm for NMR relaxometry and related experiments introduced in [1] can be extended to N D problems, by application on 2D data slices. We have proved a guarantee of successful recovery using our algorithm and demonstrated its effectiveness on simulated data and on experimental data from an olive oil sample. We find that significant subsampling can be performed while maintaining excellent fidelity of the recovered model. ACKNOWLEDGMENT This work was supported by the Intramural Research Program, National Institute on Aging, of the National Institutes of Health. AC was supported by NSF Award DMS-1402254. WC was supported by HDTRA 1-13-1-0015. R EFERENCES TABLE II O LIVE OIL T1 -D-T2 EXPERIMENTAL RESULTS 100% 25% 10 % 5% 2.5% 1% 0.0012 0.0005 0.0086 0.0019 0.0261 0.0075 0.1117 0.1573 0.0004 0.0003 0.0010 0.0007 0.0076 0.0022 0.0262 0.0097 0.2566 0.2594 1 1 1 0.66 0.12 0.0006 0.0014 0.0067 0.0580 0.2769 Relative error in peak centroid Peak 1 Peak 2 0 0 0.0005 0.0003 Relative error in peak integral Peak 1 Peak 2 0 0 Admissibility 1 Relative error in M̃ 0 X. I NCOHERENCE IN P RACTICE The incoherence for the Parseval tight frame, for the example of the simulated data, was µ = 94.17. Since r1 = r2 = 7 and |J| = 1282 , inequality (8) gives theoretical bounds 7 ≤ 2 µ ≤ 128 ≈ 2340.57. Hence, qualitatively, µ is fairly close 7 to its theoretical lower bound. This suggests that it is possible to obtain successful recovery from few measurements, as supported by theory. For the experimental data, the incoherence was 161.60, and the theoretical bounds were 4 < µ < 512. While these values of µ are not as small as desired, the 90-percentiles of ||vi ||22 maxJr1 ,r2 were 7.92 and 8.13 for the simulated and experimental data, respectively. Hence, most of the frame entries have small norm. As suggested in [1], the idea of asymptotic coherence could be used to further improve our sampling bound [21]. The bound on P in (9) depends only on µ, the rank r, and the dimension r1 and r2 , not explicitly on the size of the data |J|. We remark that while the bound on P is not useful when m1 ×m2 is small, for very large data sizes m1 ×m2 , Theorem [1] Cloninger, A., Czaja, W., Bai, R., & Basser, P. J. (Jul 2014). Solving 2D Fredholm Integral from Incomplete Measurements Using Compressive Sensing. SIAM J. on Imag. Sci., 7, 3, 1775-1798. [2] Venkataramanan, L., Song, Y.-Q., & Hurlimann, M. D. (January 2002). Solving Fredholm Integrals of the First Kind With Tensor Product Structure in 2 and 2.5 Dimensions. IEEE Trans. on Sig. Proc., 50, 1017-1026. [3] Ma, S., Goldfarb, D., & Chen, L. (Jan 2011). Fixed point and Bregman iterative methods for matrix rank minimization. Mathematical Programming, 128, 1-2. [4] Celik, H., Bouhrara, M., Reiter, D. A., Fishbein, K. W., & Spencer, R. G. (Nov 2013). Stabilization of the inverse Laplace transform of multiexponential decay through introduction of a second dimension. J. Mag. Res., 236, 2, 134-139. [5] Candes, E. J., Romberg, J., & Tao, T. (Jan 2006). Robust Uncertainty Principles: Exact Signal Reconstruction From Highly Incomplete Frequency Information. IEEE Trans. on Inf. Theory, 52, 2, 489. [6] Donoho, D. L. (Jan 2006). Compressed Sensing. IEEE Trans. on Inf. Theory, 52, 4, 1289. [7] Gandy, S., Recht, B., % Yamada, I. (2011) Tensor completion and low-n-rank tensor recovery via convex optimization. Inverse Problems. 27, 2. [8] Lustig, M., Donoho, D., & Pauly, J. M. (Dec, 2007). Sparse MRI: The application of compressed sensing for rapid MR imaging. Mag. Res. in Med., 58, 6, 1182-1195. [9] Reiter, D. A., Lin, P.-C., Fishbein, K. W., & Spencer, R. G. (Apr 2009). Multicomponent T2 relaxation analysis in cartilage. Mag. Res. Med., 61, 4, 803-809. [10] Callaghan, P. T., Arns, C. H., Galvosas, P., Hunter, M. W., Qiao, Y., & Washburn, K. E. (2007). Recent Fourier and Laplace perspectives for multidimensional NMR in porous media. Mag. Res. Imag., 25, 4, 441-444. [11] Hills, B. P. (2009). Relaxometry: Two-Dimensional Methods. Enc. of Mag. Res. [12] Liu, J., Musialski, P., Wonka, P., & Ye, J. (Jan 2013). Tensor completion for estimating missing values in visual data. IEEE Trans. on Pattern Analysis and Machine Intel., 35, 1, 208-20. [13] Candès, E. J., & Tao, T. (May 2010). The power of convex relaxation: Near-optimal matrix completion. IEEE Trans. on Inf. Theory, 56, 5, 2053-2080. [14] Fazel, M., Candes, E., Recht, B., & Parrilo, P. (Dec 2008). Compressed sensing and robust recovery of low rank matrices. Asilomar Conf. 1043-1047. [15] Gross, D. (Mar 2011). Recovering low-rank matrices from few coefficients in any basis. IEEE Trans. on Inf. Theory, 57, 3, 1548-1566. [16] Recht, B., Fazel, M., & Parrilo, P. A. (Nov 2010). Guaranteed minimum-rank solutions of linear matrix equations via nuclear norm minimization. SIAM Review, 52, 3, 471-501. [17] Chen, Y. (2015) Incoherence-Optimal Matrix Completion. IEEE Tr. on Inf. Theory. [18] Candes, E., & Recht, B. (Jun 2012). Exact matrix completion via convex optimization. Comm. of the ACM, 55, 6, 111-119. [19] Cai, J. F., Candes, E. J., & Shen, Z. (Apr 2010). A singular value thresholding algorithm for matrix completion. Siam J. on Opt., 20, 4, 1956-1982. [20] Liu, Yi-Kai. (2011). Universal low-rank matrix recovery from Pauli measurements. Adv. in Neural Inf. Proc. Sys. 1638-1646. [21] Adcock, B., Hansen, A. C., Poon, C., Roman, B. (Feb 2013) Breaking the coherence barrier: A new theory for compressed sensing. [22] Golub, G. H., & Van, L. C. F. (1996) Matrix computations. Johns Hopkins.