IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 47, NO.

5, MAY 1999

1335

Discrete Fractional Fourier Transform Based on Orthogonal Projections
Soo-Chang Pei, Senior Member, IEEE, Min-Hung Yeh, and Chien-Cheng Tseng, Member, IEEE
Abstract—The continuous fractional Fourier transform (FRFT) performs a spectrum rotation of signal in the time–frequency plane, and it becomes an important tool for time-varying signal analysis. A discrete fractional Fourier transform has been recently developed by Santhanam and McClellan, but its results do not match those of the corresponding continuous fractional Fourier transforms. In this paper, we propose a new discrete fractional Fourier transform (DFRFT). The new DFRFT has DFT Hermite eigenvectors and retains the eigenvalue-eigenfunction relation as a continous FRFT. To obtain DFT Hermite eigenvectors, two orthogonal projection methods are introduced. Thus, the new DFRFT will provide similar transform and rotational properties as those of continuous fractional Fourier transforms. Moreover, the relationship between FRFT and the proposed DFRFT has been established in the same way as the conventional DFT-tocontinuous-Fourier transform. Index Terms— Discrete Fourier transform, discrete fractional Fourier transform, Fourier transform, fractional Fourier transform.

I. INTRODUCTION HE FOURIER transform (FT) is one of the most frequently used tools in signal analysis [1]. A generalization of the Fourier transform—the fractional Fourier transform (FRFT)—has been proposed in [2] and [3] and has become a powerful tool for time-varying signal analysis. In time-varying signal analysis, it is customary to use the time-frequency plane, with two orthogonal time and frequency axes [4]. Because the successive two forward Fourier transform operations will result in the reflected version of the original signal, the FT can in the be interpreted as a rotation of signal by the angle time–frequency plane and represented as an orthogonal signal representation for sinusoidal signal. The FRFT performs a rotation of signal in the continuous time–frequency plane to any angle and serves as an orthonormal signal representation for the chirp signal. The fractional Fourier transform is also called rotational Fourier transform or angular Fourier transform in some documents. Besides being a generalization of the FT, the
Manuscript received December 2, 1996; revised February 19, 1998. This work was supported by the National Science Council, R.O.C., under Contract NSC 85-2213-E002-025. The associate editor coordinating the review of this paper and approving it for publication was Dr. Akram Aldroubi. S.-C. Pei is with the Department of Electrical Engineering, National Taiwan University, Taipei, Taiwan, R.O.C. (e-mail: pei@cc.ee.ntu.edu.tw). M.-H. Yeh is with the Department of Computer Information Science, Tamsui Oxford University College, Tamsui, Taipei, Taiwan, R.O.C. C.-C. Tseng was with the Department of Electronics Engineering, Hwa Hsia College of Technology and Commerce, Taipei, Taiwan, R.O.C. He is now with the Department of Computer and Communication Engineering, National Kaohsiung First University of Science and Technology, Taipei, Taiwan, R.O.C. Publisher Item Identifier S 1053-587X(99)03244-4.

T

FRFT has been proved to relate to other time-varying signal analysis tools, such as Wigner distribution [4], short-time Fourier transform [4], wavelet transform, and so on. The applications of the FRFT include solving differential equations [2], quantum mechanics [3], optical signal processing [5], timevariant filtering and multiplexing [5]–[8], swept-frequency filters [9], pattern recognition [10], and time–frequency signal analysis [11]–[13]. Several properties of the FRFT in signal analysis have been summarized in [9]. Many methods for realizing the FRFT have been developed, but most of them are to utilize the optical implementation [14], [15] or numerical integration. Because the FRFT is a potentially useful tool for signal processing, the direct computation of the fractional Fourier transform in digital computers has become an important issue. The ideal discrete fractional Fourier transform (DFRFT) will be a generalization of the discrete Fourier transform (DFT) that obeys the rotation rules as the continuous FRFT and provides similar results as the FRFT. In [16], a method for a numerical integration FRFT has been proposed, but the method does not obey the rotation rules, and the signal cannot be recovered from its inverse transform. In [17], Santhanam and McClellan have developed a discrete FRFT, but their method does not provide the same transforms to match those of the continuous case. In this paper, we present a new discrete fractional Fourier transform (DFRFT). This DFRFT is a generalization of the DFT and will provide similar transforms as those of the continuous case. The relationship between the DFRFT and the FRFT can also be established and discussed in detail. Moreover, the proposed DFRFT has important unitary and rotation properties. This paper is organized as follows. In Section II, the previous development of continuous and discrete fractional Fourier transforms are reviewed. The concept for developing the DFRFT to have similar results as the continuous corresponding case are described in Section III. Two acceptable solutions for the DFRFT are considered and proposed in Section III. Then, the relationships between the FRFT and the DFRFT can be established in Section IV. Finally, conclusions and discussions are made in Section V. II. PRELIMINARY A. Continuous FRFT The Fourier transform of a signal can be interpreted as a angle rotation of the signal in the time–frequency plane. The FRFT is then developed and treated as a rotation of signal

1053–587X/99$10.00 © 1999 IEEE

The conventional methods to compute the DFRFT in [17] and [20] are mainly on numerical computing the fractional power of the DFT kernel matrix. The th-order normalized Hermite function with variance is defined as (3) is the th-order Hermite polynomial [18]. . to any angles in the time–frequency plane [9]. In order to avoid ambiguity. B. MAY 1999 Fig. and ) are used to denote the rotation angles in the time–frequency plane. the two results . a rotation operation using Hermite functions has also been proposed. otherwise here is defined as [ ]. 1 shows the FRFT of the rectangular window function [ for . Furthermore. The equivalence of (1) and (2) has been proved in [3]. The discrete results shown in Fig. the Greek subscripts (for example. Imaginary part: dashed line. is the eigenfunction of the FRFT. Fractional Fourier transform of a rectangular window. a circular reflected version of the signal ( ). and the imaginary parts of the FRFT or DFRFT are indicated by dashed lines. 1. 47. the FRFT of the signal (4) (5) Equation (5) indicates that the FRFT can be interpreted as a weighting summation of Hermite functions. it will have a different phase term in the weighting coefficients. VOL. In [19]. 2 are quite different from the results of Fig.1336 IEEE TRANSACTIONS ON SIGNAL PROCESSING. and . The real parts of the FRFT or DFRFT in this paper are plotted by solid lines. elsewhere] for various angles. and the English superscripts are used to denote the fractional power values of the DFT kernel matrices in this paper. x(t) = 1 when 02  t  2. [3]. [6]. Equation where is the linear (8) indicates that the DFRFT of the signal combination of the four major angular parts: the original signal . its DFT . Because where the normalized Hermite function with unitary variance. and a circular reflected version of its DFT . [9] csc if if if is not a multiple of (1) is a multiple of is a multiple of (2) where indicates the rotation angle of transformed signal for is the th-order normalized Hermite function the FRFT. The rectangular window used . Thus. and x(t) = 0 otherwise. the (8) is the conventional DFT of signal . 1. . The weighting coefficients are obtained from multiplying the phase term and the inner product of the input signal and the corresponding Hermite function. Real part: solid line. 5. the middle is not the intermediate state of transform result for and . The transform kernel of continuous FRFT is defined as [2]. (2) provides an eigendecomposition representation of the FRFT kernel. The fractional power of the DFT kernel matrix and the rotation angles in the DFRFT essentially mean the same thing. the Hermite functions in [19] with a different scaling to those of the FRFT. The fractional power of the DFT kernel matrix [20] is calculated by (6) (6) where (7) Applying the above defined kernel to the signal is computed as DFRFT of the signal . with unit variance. and the rotation output of the signal is also the weighting summation of Hermite functions. Using the kernel of the by angle is computed as FRFT. However. 2 shows the results of this DFRFT produced by a discrete rectangular window. NO. Fig. The Old DFRFT The kernel matrix of the DFRFT is obtained by computing the fractional power for the DFT kernel matrix. Fig.

we only summarize some results for our further development of DFRFT. a method for computing the DFT eigenvectors has been introduced. . This In Table I. 1. The parameter is the index for the DFT eigenvalue .: DISCRETE FRACTIONAL FOURIER TRANSFORM BASED ON ORTHOGONAL PROJECTIONS 1337 Fig. are . Eigendecomposition of DFRFT The development of our DFRFT is based on the eigendecomposition of the DFT kernel. The output is quite different TABLE I MULTIPLICITIES OF THE EIGENVALUES OF DFT KERNEL MATRIX where matrix . is defined. Proposition 3: DFT Hermite eigenvectors should have the . . where associated continuous spread variance is the sampling intervals of signal. Old DFRFT of a rectangular window function. The eigenvalue multiplicities of DFT matrix indicate the ranks of the DFT eigensubspaces [21]. but they correspond to different eigenvalues. . and many properties of the DFT matrix eigenvalues and eigenvectors have been discussed in [20] and [21]. . . [22]. and x (n ) = 0 otherwise. Matrix commutes with the DFT kernel . we have known that the eigendecomposition of the DFT kernel matrix is not unique. a multiplicity function function is used to denote the DFT eigenvalue multiplicity for . and then. The continuous approximate form can be written as (13) Sampling by . 2. . Proof: See [21]. If the Hermite function are sampled in this way. (9) The Fourier transform of (13) can be computed as (15) .. from the continuous FRFT in Fig. In [22]. In [20]. DFT Hermite Eigenvectors III. In the discussion in the previous subsection. and matrix is defined as OF DFRFT The continuous FRFT has a Hermite function with unitary variance as its eigenfunction. x(n) = 1 when 06  n  6. Proposition 2: A matrix can be used to compute the real eigenvectors of the DFT matrix . . . where is the spread variance of the Proof: It is assumed that DFT eigenvectors. . and —and each is major eigensubspaces— . . Proposition 1: The eigenvalues of and its multiplicities are shown in Table I. Here. it satisfies the commutative property (10) will also be the eigenvectors The eigenvectors of matrix of the DFT kernel matrix . a novel matrix is introduced to compute the real and complete set of DFT eigenvectors very elegantly. . and corresponding to one of the four eigenvalues—1. . but it cannot obtain the real-valued DFT eigenvectors. we get (12) is the th-order Hermite polynomial. (13) will become (14) .PEI et al. The corresponding eigenfunction property for the DFT would be like (11) is the eigenvector of DFT corresponding to the where th-order discrete Hermite function. DEVELOPMENT A. Can the DFT have the eigenvectors with the similar shapes as the Hermite functions? These DFT eigenvectors are known as DFT Hermite eigenvectors in this paper. 1. . The eigenvectors of the DFT matrix constitute four . . —respectively. B. Proof: See [20]. .

the larger order is. The sampled Hermite functions still can be used to construct the DFRFT kernel because they have the similar shapes and good approximations to the continuous Hermite functions. both errors approach zero. because larger is . we have (20) (24) for sufficiently large . When the value of approaches infinity. Thus. Thus. For By truncating the integral interval of Fourier transform from to . The th-order continuous Hermite function should have zeros [18]. the sampling interval we will prove that even such samplings of Hermite functions can only have approximate Hermite eigenvectors of the DFT. The small aliasing will occur while is closer to . VOL.1338 IEEE TRANSACTIONS ON SIGNAL PROCESSING. and the samplings in Proposition 3 cannot guarantee that the numbers of the sign changes in this sampled th-order Hermite function are also . the better approximation (17) is. approximated by can be From Proposition 3.. Proof: Here. Thus. it is clear that there are two approximation errors in (17). the decay rate the degree of Hermite polynomial is proportional to for of the Hermite function sufficiently large . we can find that the DFT Hermite eigenvectors should have associate continuous spread variance . the slower decay rate the Hermite function will have. and our goal is to adjust the variance have the same eigenvector shape as itself. This implies that the truncation error in (19) is larger for high-order . The condition of for the two shapes of (14) and (16) being the same becomes This approximation is also valid because is is very large. The sequence obtained from (12) is also equal to the sampling of the Hermite function with the unit variance by . Combining (19) and (20) very small when . frequency spectrum that DFT can cover is the continuous spectrum spacing (or resolution) of the DFT is . One is the truncation error in (19). by replacing the continuous integral with the numerical integral. we only prove the case that odd. the proof can be easily derived in the same way.e. From the above proof. The variable in (15) can be replaced by equal to to have its discrete form. these functions are not bandlimited. In Propositions 4 and 5. . and the decay rate of Gauss function very fast. is obtained by samProposition 4: If the sequence pling the unit variance Hermite function with the sampling . but this will not influence the development of the DFRFT. Next. we obtain and letting This expression is valid for any both sides. we have the approximation expression (19) is wide This approximation is valid because is for large . NO. the is. when order becomes large. we have . Next. MAY 1999 The sampling theorem [23] tells us the continuous scope of . 5. for (23) for then it can be shown that the DFT of the . However. then it can be shown that interval If is even (17) If is odd (18) for sufficiently large . taking at (21) Equation (14) becomes The proof is completed. Thus. the approximation in (17) becomes worse. (16) adjusts the scaling of Hermite funcThe variance value to make the DFT tion. Therefore. is even. defined in the range Proposition 5: If the sequence is obtained by shifting Hermite Gauss samples in the following way: If is even for (22) for and if is od. and the other is the numerical error in (20). 47. . i.

it has been proved that the samplings of Hermite functions can have approximate DFT eigenvectors. Unfortunately. This assignment rule matches the multiplicities of the eigenvalues of the DFT kernel matrix in Table I. Combining . In Propositions 4 and 5. therefore. The While is given by DFT of the sequence DFT TABLE II EIGENVALUES ASSIGNMENT RULE OF DFRFT KERNEL MATRIX (25) Using the equality second term on the right side of (25) becomes . and is the eigenvector of matrix where . .PEI et al. The normalized vectors for the samplings of Hermite functions are defined as (31) . It is because the approximation error of the low DFT Hermite eigenvectors are small. the proof can also be easily derived. In (32). we know that there are several DFRFT kernel matrices that can obey the rotation properties. Using the . we also prove the case that is even. In the case of continuous FRFT.: DISCRETE FRACTIONAL FOURIER TRANSFORM BASED ON ORTHOGONAL PROJECTIONS 1339 Proof: Here. Because of the above ambiguity. This indicates that there exists root ambiguity in deciding the fractional power of eigenvalues. • Ambiguity in Deciding the Eigenvectors of the DFT Kernel Matrix: The DFT eigenvectors constitute four major eigensubspaces. the terms of the Hermite functions are summed up from order zero to infinity. The idea for developing our DFRFT is to find the discrete form for (2). The selections of the DFT Hermite eigenvectors are from low to high orders. • Ambiguity in Deciding the Fractional Powers of Eigenvalues: We know that the square root of unity are 1 and 1 from elementary mathematics. In addition. the choices for the DFT eigenvectors to construct the DFRFT kernel are multiple and not unique. Newly Developed DFRFT DFT (27) (28) is limited in the range where equality rewritten as DFT (29) is limited in the range where (28) and (29). In order to retain the eigenfunction property in (11). is odd. In addition. (26) Substituting (26) into (25) and using Proposition 4. eigenvectors for the However. we obtain C. the eigenvectors can be used as bases for individual DFT eigensubspaces. there exist two types of ambiguity in deciding the fractional power of the DFT kernel matrix. we should not expect that a finite vector can express the oscillation behavior of the very high-order Hermite function very well. (27) can be The fractional power of matrix can be calculated from its eigendecomposition and the powers of eigenvalues. will be a DFT Hermite eigenvector. we obtain DFT for for (30) The proof is completed. Table II shows the eigenvalues assignment rules for the DFRFT. the DFT is computed from the eigenvectors of Hermite eigenvector matrix in the same DFT eigensubspace. the unit variance Hermite functions are in the following sampled with a period of discussions. we can compute in its DFT eigensubspace to obtain a the projections of Hermite-like DFT eigenvector (32) mod . the Because matrix can have complete real orthogonal DFT eigenvectors. only DFT Hermite eigenvectors can be added. for the discrete case.

NO. and for even . 47. Therefore. for and the eigenvalues of the DFRFT are . The DFT Hermite eigenvectors in each DFT eigensubspace can be orthogonalized by the Gram–Schmidt method. and for is even. In addition. vector orthogonalization is required for the DFT Hermite eigenvectors obtained from (32). shown at . The Gram–Schmidt Algorithm (GSA): The Gram–Schmidt method [29] is a well-known orthogonalization approach for vectors. . . the transform kernel for the rotation angle can be computed as in (34a). This means that is replaced by in (34). Unfortunately. It must be mentioned that the last eigenvalue and should be assigned to the is not assigned to to match the eigenvalue multiplicities value of the DFT matrix in Table I. the eigenvalue–eigenfunction relation is retained in defining the DFRFT. . Therefore. is defined as follows: where For odd . is the normalized eigenvector corresponding to the th-order Hermite function. the eigenvectors obtained from matrix are directly considered to be the discrete Hermite functions. .1340 IEEE TRANSACTIONS ON SIGNAL PROCESSING. The role of matrix in (32) is just as a tool to find a complete set of real and orthogonal DFT eigenvectors. An eigenbased method for computing the DFRFT to have similar continuous results has been proposed by us in [24]–[27]. orthogonalized Hermite eigenvector. It is easy to verify that the angle rotation property of the DFRFT can be preserved only while the DFT eigenvectors are orthogonal.. nonorthogonalized Hermite eigenvector. Equation (32) provides a method for finding DFT Hermite eigenvectors. Two methods for vector orthogonalization are proposed in the following of this paper. the eigenvectors obtained from the matrix are just discrete Mathieu functions [21]. However. VOL. So the DFT Hermite eigenvectors can be orthogonalized for every eigensubspace individually to obtain orthogonal eigenvectors in the whole eigenspace of DFT. The symbol notation in developing the two algorithms are as follows: continuous Hermite samples vector. the convergence for the eigenvectors obtained from matrix are not so fast for the high-order Hermite functions by the LMS error criterion. In [25]. (34a) . . Although the Mathieu functions can converge to a Hermite function [28]. This is because the eigenvectors located in different eigensubspaces will be always orthogonal. It is easy to show that the eigenvectors located in different eigensubspaces will be orthogonal. . whereas is odd. MAY 1999 The transform kernel of the DFRFT can be defined as (33) for (odd) (34) for (even) where . . The DFT eigenvectors are orthogonalized for each eigensubspace.. . Calculate the continuous samples of Hermite functions: Compute the eigenvectors of matrix : Using equation (32) to compute Hermite eigenvectors by projections: for to for to end end . . the DFT Hermite eigenvectors obtained from (32) cannot constitute an orthogonal basis for DFT eigenspace. 5. where the eigenvector corresponding to the th continuous Hermite function. Example 1— : The rotation angle is equal to . is the bottom of the page. . Such a method is called the method in this paper.

respectively. The continuous FRFT is an orthonormal signal decomposition for chirp signals [9]. Implementation of the New DFRFT As in the case of DFT frequency domain. 7. The sampled discrete data then becomes . Frobenius norm of (35) samples of Hermite functions are plotted for . satisfy Algorithm Calculate the continuous samples of Hermite functions: Compute the eigenvectors of : Using equation (32) to compute Hermite eigenvectors by projections: for to Compute the SVD of . Figs. Orthogonal Procrustes Algorithm (OPA): A traditional mathematical problem known as the orthogonal Procrustes Algorithm can be used to find the least Frobenius norm [30. In Fig.: DISCRETE FRACTIONAL FOURIER TRANSFORM BASED ON ORTHOGONAL PROJECTIONS 1341 Fig. 8–10. [ which is the same as the signal used in Fig. . the last half of the indices in the DFRFT must also be treated as the negative end The GSA minimizes the errors between the samples of Hermite functions and orthogonal DFT Hermite eigenvectors from low to high orders. 9 and 10. 3. 4–6 show the DFRFT of the rectangular window function for various angles for method. Example 3: The DFRFT by the GSA for a chirp signal is computed in this example.PEI et al. and OPA. 2. D. p. Based on the unitary property in the DFRFT and the transform results shown in Figs. 3. Because the . The error vectors between samples of Hermite functions and DFT Hermite eigenvectors are defined as (37) Example 2: In this example. the solution will also . and the OPA minimizes the total errors between the samples of Hermite functions and orthogonal DFT Hermite eigenvectors. for minimize minimize where matrix. The is equal to 4/13. Properties of DFRFT The properties of the DFRFT are shown in Table III. ) are plotted Transform results of the impulse signal ( in Figs. . GSA. respectively. E. A more detailed theory and algorithm can be found in [7]. 1. Therefore. and will can be found by be our solution. ( the DFRFT can be used for the chirp signal and chirp rate detection. 8–10 for the method. The corresponding samples of the continuous FRFT for the impulse signal are plotted in Fig. we use the DFRFT to deal with the rectangular window function shown in Fig. The chirp signal used here is equal . where . . otherwise. The minimizing calculating the singular value decomposition (SVD) of . In Fig. ]. The norm of the error vectors between the DFRFT and samples of the FRFT are also shown in the titles of Figs. the norms of error vectors between the computed DFT Hermite eigenvectors and . and OPA. it can be to found that the transform results change from a chirp signal ) to an impulse-like function ( ). We can formulate our problem as the OPA to find the least Frobenius norm between the samples of Hermite functions and orthogonal DFT Hermite eigenvectors. 11. we can find that the proposed DFRFT provides a similar orthonormal signal decomposition for discrete chirplike signals. Norms of error vectors between DFT Hermite eigenvectors and samples of Hermite functions. is the subject to (36) the . GSA. and the number of points sampling interval is equal to 73. 582] for the two given spaces.

the following method for implementing the DFRFT can be applied: (38) (39) frequency. computing the DFRFT is . Then. 47. If the rotational angles are adjusting. k = 032 1 1 1 32. closer output to the continuous FRFT than Fig. 6. DFRFT by S method of a rectangular window function. 1. When = 3=7. and x (n ) = 0 otherwise. DFRFT by GSA of a chirp signal. output to the continuous FRFT than Fig. 5. This figure has a closer Fig. ( and the rotation angle are When the number of points determined. the computation of the DFRFT can be implemented only by a transform kernel matrix multiplication. The complexity of . 5. x(n) = 1 when 06  n  6. identity transform) and any angle transform domains. and it is the same as in the DFT case. DFRFT by OPA method of a rectangular window function. x (n ) = 1 when 06  n  6. MAY 1999 Fig. 7. 4. and the eigenvalues of DFRFT are also determined. This concept is also applied in the time domain . ej 20:1141k . 4. an impulse-like output is obtained. the continuous FRFT in Fig. the DFT Hermite eigenvectors can be computed a priori. 4. x (n ) = 1 when 06  n  6. and x(n) = 0 otherwise. VOL. This figure has a Fig. NO. The output is close to Fig. and x (n ) = 0 otherwise.1342 IEEE TRANSACTIONS ON SIGNAL PROCESSING. DFRFT by GSA method of a rectangular window function.

Proof: Let be any DFT eigenvector located in the . Fig. x(0) = 1. and the output becomes a constant. DFRFT by OPA of an impulse function. Applying the transform eigensubspace kernel defined in (6) to the eigenvector . It can be seen that chirp-like outputs are obtained while the angle = =2. A rigorous discussion for the mismatches of [17] has been presented in [16]. and x(n) = 0 otherwise. When = =2. the diagonal matrix F. Discussion The method in [17] obeys the rotation properties. 10. 9. It can be seen that chirp-like outputs are obtained while the angle = =2. It can be seen that chirp-like outputs are obtained while the angle 6= =2. DFRFT by S method of an impulse function. . and x(n) = 0 otherwise. the inner products of signal and eigenvectors. 6 Fig. Proposition 6: The method in (6) assigns all the eigenvectors of the DFRFT matrix to only four different eigenvalues. x(0) = 1. This is the major reason for the mismatches in [17]. When = =2. and x(n) = 0 otherwise. and they can be computed in advance.: DISCRETE FRACTIONAL FOURIER TRANSFORM BASED ON ORTHOGONAL PROJECTIONS 1343 Fig. we can obtain is also a DFRFT eigenvector. but it cannot have similar results as in the continuous case. The coefficients ’s are in (33). When = =2. and the output becomes a constant. DFRFT by GSA of an impulse function. . and the value is the eigenvalue for this eigenvector . Since any eigenvector in Therefore.PEI et al. we stress the major reason for this mismatch. Here. the DFRFT reduces to DFT. If the rotation angle is changed. 8. the DFRFT reduces to DFT. 6 (40) The definitions of matrix and matrix are the same as those . x(0) = 1. the DFRFT reduces to DFT. only should be recomputed. and the output becomes a constant.

which satisfies the eigenvalue and eigenfunction relationship as a continuous FRFT kernel.1344 IEEE TRANSACTIONS ON SIGNAL PROCESSING. NO. . the periods in the transform results change for different rotation angles. we will review and understand the transform range and resolution for the is the conventional DFT.b)-element in DFRFT kernel matrix F [vn ] = e0j n vn where vn is Hermite-like function N F [(k)]  2 10j2cot ej cot  If x is even. the method assigns the DFRFT eigenvalues to only four values. the number of DFT Hermite eigenvectors is only finite. and there are some approximation errors in the DFT Hermite approaches infinity. X is odd. sampling interval for the original continuous signal. is the total recorded signal duration and is signal. On the other hand. the signal and transform results in [31] are discrete and periodic. the transform range and resolution of the DFT have been well discussed [23]. an alternative DFRFT has recently been proposed. our DFRFT can be used for any rotation angle and can provide very similar results as continuous cases. the proposed DFRFT has assigned a different eigenvalue for each eigenvector. However. and transform range in the time domain as well. In the following discussion. and x(t) = 0 otherwise. the DFRFT can be used to give the similar continuous transform results within the accuracy of the discrete finite vector approximation. The DFRFT developed in this paper is based on the mimicking of the eigenvalue–eigenfunction relationship that the continuous counterpart of this transform has with the normalized unit variance Hermite functions. In the DFRFT of [17]. it has been proved that the spread variance . and the frequency resolution is [23]. Here. In the conventional DFT analysis. 47. If x is odd. It should be noted that the number of eigenfunctions for the continuous FRFT in (2) is infinite. and the rotation angle in [31] is valid for a certain discrete set of rotation angles. there will be four different eigenvalues for all the eigenvectors. we will of the DFT Hermite eigenvector is compute the FRFT for a Hermite function with any variance . While errors of the DFT Hermite eigenvectors will be reduced. ( one of the four eigensubspace has the same eigenvalue. b) is the (a. we will discuss the transform range and resolution for the DFRFT (see Table IV). To begin with. Thus. Then. Transform Range and Resolution of DFRFT In this subsection. Moreover. Samples of continuous FRFT of an impulse function. In Proposition 3. TABLE III PROPERTIES OF DFRFT Unitary Angle additivity Time Inversion Periodicity Symmetry Eigenfunction Impulse Transform Parity F )3 = (F )01 = F0 F F =F F x(0n) = X (0n) F =F F (a. 11. a) where F (a. RELATIONSHIP BETWEEN FRFT AND DFRFT In this section. the approximation eigenvectors. It is based on retaining the property of the DFT that a sampled periodic function transforms into a periodic function. IV. it is also is the number of points of the discrete the time resolution. VOL. (41) is the eigenvector corresponding to the th-order where Hermite function. 5. Therefore. b) = F (b. X is even. A. and more DFT Hermite eigenvectors are used to compute the DFRFT. The development of the new DFRFT is based on the same idea as (2). we will establish the relationship between the FRFT and the DFRFT. MAY 1999 Fig. The the overall frequency range that the DFT can represent is equal to . In [31]. x(0) = 1.

12. • The lengths of time and frequency ranges are equal ). will be equal to 1 for any value of . The sampled spread variance of the DFRFT is vector in (12) can also be an approximate eigenvector in the fractional Fourier domain (angle ). as drawn in Fig. we can substitute the variance in (43) by to get the spread variance of the DFRFT in the fractional Fourier domain (angle ) (48) Both (47) and (48) indicate the corresponding variance of DFT Hermite eigenvectors in the fractional Fourier domain. therefore. whereas This means that the DFRFT is implemented in a circular rotation for the discrete case. (46) Thus. It is easy to check some special cases for (43): and . 12. The overall transform range of the DFRFT can cover is equal to (45) Proof: From Proposition 3. only the last term can affect the envelope . ( • The length of the time range is longer than that of the ).PEI et al. is used to denote the resolution of the FRFT. The new variance is com- (43) Proof: This proposition can be easily proved by computing the FRFT of the normalized Hermite functions. In (42).: DISCRETE FRACTIONAL FOURIER TRANSFORM BASED ON ORTHOGONAL PROJECTIONS 1345 TABLE IV TRANSFORM RANGE AND RESOLUTION OF DFRFT Proposition 7: The continuous FRFT of a Hermite function with variance is equal to Moreover. • Circular Rotation: (51) . The symbol and are the two special cases. Angle actual angle for the elliptical rotation. but it actually performs elliptical rotation in the continuous time–frequency plane. For simplification of analysis. and is the where number of points for the discrete signal. but the DFRFT performs elliptical rotation in the continuous time–frequency plane while . Elliptical Rotation versus Circular Rotation The FRFT performs circular rotation in the continuous time–frequency plane. Proposition 8: The resolution of the DFRFT with angular is equal to parameter (44) is the sampling interval of the signal. 12. it can be found that the signal rotation of the DFRFT is an elliptical rotation in . If . of the FRFT of From Propositions 4 and 5. it has been known that the . In this subsection. the resolution of the FRFT can be obtained as (49) (42) where puted as . frequency ( The rotation concept of the DFRFT in the continuous is the time–frequency plane is plotted in Fig. frequency ( • The length of the time range is shorter than that of the ). the continuous time–frequency plane. we will define circular and elliptical rotations clearly and establish the relationship between circular and elliptical rotations. the DFRFT performs independent circular rotation in the discrete notation. From (45) and Fig. it has been shown that there are approximation errors in the DFT Hermite eigenvectors. B. the spread variance of the DFRFT for angle is equal to (47) The overall transform range the DFRFT can cover is equal to (50) In (34). Three cases can be realized for the signal rotation of the DFRFT in the continuous time–frequency plane. the approximation errors of the DFT Hermite eigenvectors are ignored in the following discussion.

The transform kernels of the circular and elliptical rotations defined above are infinite numbers of sums for Hermite eigenvectors. and where is defined as in (43). is the elliptical rotation . 47. • Elliptical Rotation: where (54) (55) (56) spread variance changes in these two schemes. (59) . and the post-phase compensa- is the transform kernel of elliptical rotation. Results of Example 4 showing that the error norms between DFRFT and continuous FRFT samples are greatly reduced by angle modification and post-phase compensation. (52) (53) is the transform kernel of circular rotation. If the Hermite eigenvectors in the circular rotation are from order 0 to infinity. 5. Three different cases of rotation concept for DFRFT in continuous time-frequency plane. is the circular rotation of signal .1346 IEEE TRANSACTIONS ON SIGNAL PROCESSING. VOL. NO. 12. 13. (a) (b) (c) Fig. the circular rotation is just the FRFT of the signal. Proposition 9: The circular rotation of signal can be implemented by an angle modification of the elliptical rotation and post-multiplying a phase compensation factor. The critical point for which we care is the (58) can be written as the weighted Proof: The signal sum of the normalized Hermite functions. While . the elliptical rotation will of signal become the circular rotation. (57) where tion factor is equal to . MAY 1999 Fig.

Example 4: In this example. Two vector orthogonalization processes for the DFT Hermite eigenvectors are accordingly proposed in this paper: One is GSA. The operation in the DFRFT performs elliptical rotation in the continuous time–frequency plane for the case . 13(c) very well. It must be noted that the postphase where compensation factors in (58) and (62) are different. With the help of the commutative matrix . 13(a) is quite different from Fig. If variance of the continuous-time counterpart of the eigenvector will approximate to unity. it will become evaluated at the points (60) (61) can be Equation (61) indicates that the FRFT with angle implemented by a DFRFT with angle . Step 2) Calculate the -point DFRFT of the signal with the angular parameter . The new transform retains the eigenvalue–eigenfunction relationship using the sampled version of the normalized unit-variance Hermite functions that the continuous FRFT has with the unit variance Hermite functions. we only compute the results of by the GSA. the DFRFT in Example 2 is almost a circular rotation. Therefore. which are equal to continuous counterpart of the variance in this example is . The GSA minimizes the errors between the samples of the Hermite functions and the orthogonal DFT Hermite eigenvectors from low to high orders. In this case. 13(b) is the DFRFT with angular parameter modification and postphase compensation. the sampling . The discrete post-phase compensation resolution is still factor for the DFRFT is the length of time range ( ) will almost equal the length ). In Example 2. the modification of the angle and postphase compensation discussed above are very critical. (62) . Fig. the following steps must be used to make the results of discrete cases match those of the continuous case. for the indices . While the variance is far from unity. the elliptical rotation can be implemented by circular rotation. 13(b) match the corresponding continuous FRFT cases shown in Fig. and the transform . The DFT Hermite eigenvectors can be calculated by the projection of samples of the unit variance Hermite functions in the DFT eigensubspaces through the help of the eigenvectors of the matrix . The correis equal to 1. CONCLUSIONS The development of this DFRFT is based on the eigendecomposition of the DFT kernel matrix . and Fig. such DFT Hermite eigenvectors cannot form an orthogonal basis for the DFT eigenspaces. the complete real and orthonormal eigenvectors of the DFT kernel matrix can be computed. Step 3) Multiply the result obtained from Step 2 by the post-phase compensation factor shown in (62). The in (58) is replaced by in (62) for preserving variable . V. If the values of (57) are . Fig. Using Proposition 9. and the other is OPA. and the effects of angle modification and postphase compensation will be very small. However. Here. Compute the FRFT of (circular rotation) for the both sides of (59). However. The rotation will almost be of the frequency range ( a circular rotation. we again deal with the rectangular window shown in Example 2. the the unitary property in the DFRFT. whereas the OPA minimizes the total errors between the samples of the Hermite functions and the orthogonal DFT Hermite eigenvectors. we directly compute the DFRFT without angle modification and postphase compensation. and the number of points in the interval is still signal becomes 37. Fig. 13(c) due to the elliptical rotation and not in the circular rotation in the continuous time–frequency plane.PEI et al. 13(c) shows the sample values of the continuous FRFT . Step 1) Compute the modification angle . therefore. 13(a) shows the DFRFT with angular parameter for the original signal. The results shown in Fig.: DISCRETE FRACTIONAL FOURIER TRANSFORM BASED ON ORTHOGONAL PROJECTIONS 1347 where . The rotation angle .0488 in Example sponding variance 2. . The length of the time range is almost equal to the length of the frequency range. where .

1953. [15] . Ozaktas. vol. NJ: Prentice-Hall..” J. W. 474–478.” J. M. Englewood Cliffs. [29] S. time–frequency analysis. 47. Feb.” in Proc. He then received the M. Taiwan. M. vol. T. Taipei. “Generalized self-Fourier functions. Namias. Chunghwa Telecom Company.. A. Y. May 1996. Rev. 45.C. [17] B. M. Onural. Hwa Hsia College of Technology and Commerce. [7] R. June 1996. degree. H. “The discrete fractional Fourier transform. vol. “The fractional Fourier transform and time-frequency representation.O. [30] G. IEEE Int. Hsinchu. “Improved discrete fractional Fourier transform.. [3] V. IEEE Int. 15–16. 10. vol. B. 33.” IEEE Trans. Time-Scale Anal. Taiwan. 3. “Digital computation of the fractional Fourier transform. M. and A. and L. and L. If the DFRFT performs a circular rotation of the signal in the . Opt. 1995. The Fourier Transform and Its Applications. if the DFRFT becomes an elliptical rotation in the continuous time–frequency plane. “Convolution and filtering in fractional Fourier domain. Soc. Mendlovic. vol. Signal Processing. 1992. The complexity for . He is currently an Assistant Professor with the Department of Computer and Communication Engineering. He is currently an Assistant Professor with the Department of Computer Information Science. 25–31. all in electrical engineering..-C. Yeh. H. vol. Arikan.” IEEE Trans. From 1995 to 1997. [28] P. and multiplexing in fractional Fourier domains and their relationship to chirp and wavelet transforms. Dr. C. W. “Fractional Wigner distribution function. pp.” IMA J. vol. 547–559.S. 1996. A.E. NTU. A. Mendlovic and H.. Apr. A. This DFRFT provides a method for implementing the discrete FRFT. [22] G. Lohmann. N. “Fractional Fourier transformations and their optical implementation—I. in 1970 and the M. M.” J.” J.” J. July 15. Opt. 1997. B. MD: Johns Hopkins Univ. J. 303–309. pp.C.-H. On the other hand. optical information processing. 665–675. Matrix Computations.. 1972. Opt. [8] A. Bozdagi.. 5. 25. Kutay. H. degrees from the University of California. 1875–1881. Ozaktas. Tseng. K. 119–124. REFERENCES [1] R. Nov. “On Namias’ fractional Fourier transforms. Taipei. Kutay. Ozaktas. . Feb.. Chien-Cheng Tseng (S’90–M’95) was born in Taipei. and it is an important tool for signal processing. pp.. 21–67. he was an Associate Research Engineer at Telecommunication Laboratories. p. Arikan. [4] F. June 1994. [19] B. 1987. 10. Steiglitz. Bracewell. and G. A. [23] A.” IEEE Trans. R. “Eigenvalue and eigenvector decomposition of the discrete Fourier transform. [10] D. W. McClellan. 1995. Interscience. C.” in Proc.E. Santhanam and J. degree in computer science and information engineering in 1992 and the Ph. Parks. 9. [6] H. Pei is a Member of Eta Kappa Nu and the Optical Society of America. from 1981 to 1983. Bourdeaux-Bartels. vol. H. . He received the B. Dragonman. 67. both from the National Taiwan University. His research interests include digital signal processing. Taipei. “Linear and quadratic timefrequency signal representations. 1994. vol. His main research interests are in fractional Fourier transforms. ASSP-30. Inst. Ltd. pp. pp. [12] D. 199–203.” Appl. “The discrete rotational Fourier transform. Taiwan. vol. 1994. Amer. “Two dimensional discrete fractional Fourier transform.. pp.S. in 1964. Taipei. pp.O. 11. G.” Signal Process. and D. “The fractional order Fourier transform and its application to quantum mechanics. 1994. Friedberg.” in Proc. Ozaktas. 1993. New York: McGraw-Hill. pp. vol. 34. pp. pp. 1994. and wavelets. Amer. vol. 99–108. M. 40–41. 1998. Opt. in 1990 and 1995. Sept. 1982. 2141–2150. Phys. degree in electrical engineering in 1997.. pattern recognition. Erkaya.. 1989. M. and M. F. H. pp. and electronic commerce. vol. degree in computer engineering from the National Chiao-Tung University. [27] S.” IEEE Signal Processing Mag. 66–74. vol. NO. 1959. His research interests include digital signal processing. pp. Mar. Signal Processing. he was a Research Assistant at UCSB. [14] D. . Soc. Signal Processing. Apr. Amer. Barshan. [5] H. Opt. V. [16] H. Feb. W. [18] G. Pei and M.C. vol. vol. Oppenheim. H. Sansone. Taiwan. Symp. Gori. 1996. Mendlovic.1348 IEEE TRANSACTIONS ON SIGNAL PROCESSING. 1989. Mendlovic. pp. in 1972 and 1975. He received the B. O. 44. Kutay. and Ph. and R. 3084–3091. Tatung Institute of Technology. Taipei. respectively. vol. [2] A. and Ph. NJ: Prentice-Hall. and M. Linear Algebra. 1989. 42. C. Tamsui. pp. “Discrete fractional Fourier transform. R. “Fractional Fourier transformations and their optical implementation—II. H. J. B. Methods of Theoretical Physics. 42. pp. Appl. pp. W. Soc. From 1971 to 1975.D. From 1997 to 1998. Dickinson and K. 7599–7602. Press. Cincotti. [21] J. Taoyuan. Discrete-Time Signal Processing. AU-20. 1047–1049. [13] H. Opt. vol. vol.O. filtering. 1. pp. Acoust. N..S. He was an Engineering Officer in the Chinese Navy Shipyard from 1970 to 1971. E.. vol. H. Math. pp.. he was an Assistant Professor of the Department of Electronics Engineering. Soo-Chang Pei (SM’89) was born in Soo-Auo. M. A. Soffer. A. “Fractional correlation. Kerr. National Kaohsiung First University of Science and Technology. in 1949. and J. Yeh. Amer. Dorsch. 1996. from the Tatung Institute of Technology. [20] B. “Relationships between the Radon–Wigner and fractional Fourier transforms. R. Amer. Lohmann. The DFRFT proposed in this paper not only supplies the similar transforms to match with those of the continuous case but also preserves the rotation properties. Orthogonal Functions. Jan. 1986.” Opt. MAY 1999 Furthermore. [11] H. W. which is the same as that implementing the DFRFT is of the DFT. 205–207.” IEEE Trans. Spence. Speech. 1998. 6th IEEE DSP Workshop. Ozaktas. He was a Professor and Chairman with the Electrical Engineering Department. Aademir. 1980. Taipei. degree from the National Taiwan University (NTU).” IEEE Trans. vol. Circuits Syst. He received the B. M. Soc. Golub and C. image processing. D. Weisburn. Circuits Syst. Barshan. New York: McGraw-Hill. G. “Chirp filtering in the fractional Fourier domain. Signal Processing. [25] 22. Audio Electroacoust. Santarsiero.S. M. New York.S. pp. A. Ozaktas.” J. 1191–1194. A. Applicat. “Eigenvectors and functions of the discrete Fourier transform. vol. McBride and F. 1798–1801. in 1987.” Opt. Mendlovic. Mar. [31] O. Math. Soc. pp. and D. F. Ozaktas. M. VOL. 1993. he is Professor and Chairman of the Electrical Engineering Department. “Discrete fractional Hartley and Fourier transforms. Santa Barbara (UCSB). 2nd ed. Tamsui Oxford University College. 11. Baltimore. in 1988. 1996.. Taipei. pp. 994–998. Taiwan. Parks. in 1965. Symp. Ozaktas. F. 39. Presently. respectively..” J.” IEEE Trans.E. Opt. C. pp. 2522–2531. pp. McClellan and T.” Signal [26] Process. “Fractional Fourier domains. pp. Bitran. Feschbach. Min-Hung Yeh was born in Taipei. Englewood Cliffs. 1994. Insel. and laser holography. Oct. [9] L.” IEEE Signal Processing Lett. Morse and H. and O. A.D. Lett. [24] S. 25. 13. with honors.E. 241–265. the relationship between the FRFT and the .D. Almeida. and the M.. pp. degrees from the National Taiwan University. 159–175. 1992. A. 536–539. Pei. 46. Van Loan. “Convolution. Hlawatsch and G. Lohmann and B. 1416.. An angle modification and a postphase compensation in the DFRFT for elliptical rotation are required to match the results that are similar for the continuous FRFT. Shenoy. Taipei. “Effect of fractional Fourier transformation on time-frequency distributions belonging to the Cohen class. Shyu. DFRFT can be established as follows.. Time–Freq. time–frequency plane.S. “Separation of transient signals. M. II. H.” Appl.

Sign up to vote on this title
UsefulNot useful