Discrete Fourier transform (DFT) class matrix fast approximation method based on integer sparse matrix factorization

Through integer sparse matrix factorization and double butterfly hybrid basis optimization model, the DFT matrix is ​​approximately decomposed into integer matrix continuous multiplication form, solving the problems of DFT calculation complexity and error, realizing the reduction of hardware complexity and error, and is suitable for High-dimensional wireless communication scenario.

CN118939928AInactive Publication Date: 2024-11-12NANTONG UNIV
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411057079.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-02
Publication Date
2024-11-12
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing DFT computing has high demands in hardware complexity and algorithm complexity, especially under the development of wireless communication technology, which has led to an increase in chip resource overhead, making it difficult to effectively reduce computing complexity and errors.

Method used

The DFT matrix fast approximation method based on integer sparse matrix factorization is adopted. The DFT matrix with N=2t with double butterfly mixed basis optimization model is approximately decomposed into integer matrix continuous multiplication form, combining the stage search optimization method and butterfly optimization algorithm to find Optimal parameters to reduce hardware complexity and error.

Benefits of technology

It successfully reduces hardware complexity and calculation errors, realizes more efficient DFT matrix decomposition, and is suitable for high-dimensional DFT matrix calculations, reducing hardware overhead and errors.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118939928A_ABST
    Figure CN118939928A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of DFT (Discrete Fourier Transform) class matrix fast approximation, in particular to a DFT class matrix fast approximation method based on integer sparse matrix factorization, which is used for researching a low-complexity calculation scheme of a DFT class matrix under different constraint conditions and carrying out fitting approximation by using a series of sparse integer matrix successive multiplication. Based on a butterfly idea, a double-butterfly mixed base optimization model is established, decoupling design is carried out for multi-parameter optimization, according to a multi-stage search optimization method, the first stage aims at reducing hardware complexity, and the second stage aims at improving fitting precision, namely, minimizing errors. Under different constraint conditions, a double-butterfly mixed base optimization model is established to approximately decompose a DFT matrix FN of which N is equal to 2t and t is equal to 1, 2, 3,... into integer matrixes A1, A2,... AK in a continuous multiplication form so as to reduce the calculation complexity, and then optimal parameters are found out under the conditions of popularizing to a DEF-like matrix and limiting the precision tolerance, so that the accuracy of the DFT matrix FN is improved, and the accuracy of the DFT matrix FN is improved. The hardware complexity can be better reduced, and errors can be better reduced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of DFT-like matrix fast approximation, and in particular to a DFT-like matrix fast approximation method based on integer sparse matrix factorization. Background Art

[0002] Discrete Fourier Transform (DFT) plays a very important role in many fields such as digital signal processing, communication signal processing, channel estimation, etc. Therefore, the calculation of DFT plays a vital role. Signal correlation, filtering, spectrum estimation, etc. can all be achieved through DFT. However, the hardware complexity of DFT calculation is related to its algorithm complexity and the range of data elements. The higher the algorithm complexity and the larger the data value range, the greater the hardware complexity. It can be seen from the definition of DFT that N points of DFT need N 2 When N is large, the amount of calculation is quite large.

[0003] In actual products, in order to quickly implement DFT and reduce computational complexity, the Fast Fourier Transform (FFT) algorithm is usually used. Based on the divide-and-conquer strategy and the idea of ​​butterfly computing, FFT uses the properties of DFT transform to significantly reduce the computational complexity of DFT. However, although chips have made great breakthroughs in various fields, there are still some limitations. With the development of wireless communication technology, the number of antenna arrays, channels and communication bandwidth continues to increase, and the demand for FFT is also increasing, resulting in greater hardware overhead for implementing FFT on dedicated chips. Summary of the invention

[0004] The purpose of the present invention is to further reduce the chip resource overhead, and a fast DFT-like matrix approximation method based on integer sparse matrix factorization is proposed. Under different constraints, a double butterfly mixed basis optimization model is established to set N=2 t , DFT matrix F for t = 1, 2, 3, ... N Approximate decomposition into integer matrices A1, A2, ..., A K The form of continuous multiplication is used to reduce the computational complexity, and then the optimal parameters are found when it is extended to the DEF-like matrix and the accuracy tolerance is limited, which can better reduce both the hardware complexity and the error.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A fast approximation method for DFT-like matrices based on integer sparse matrix factorization is proposed. A double-disc mixed basis optimization model is used to optimize matrix decomposition: based on the stage search optimization method, the matrix is ​​first optimized using a mixed basis, and then the real-valued matrix scaling factor β is optimized. Under different constraints, a double-butterfly mixed basis optimization model is established to set N=2 t , DFT matrix F for t = 1, 2, 3, ... N Approximately decompose into the form of integer matrix multiplication to reduce the computational complexity, and then find the optimal parameters when generalizing to the DEF-like matrix and limiting the accuracy tolerance, which can better reduce both hardware complexity and error; including the following steps:

[0007] F N is an N-dimensional DFT matrix in the following form:

[0008]

[0009] Where j is the imaginary unit;

[0010] Consider the following two constraints:

[0011] Constraint 1: Limit Each matrix A k Each row of has at most 2 non-zero elements;

[0012] Constraint 2: Limit Each matrix A k Meet the following requirements:

[0013]

[0014] Among them, A k [l, m] represents the matrix A k The element at row l and column m, A k The range of the real and imaginary parts of the elements in , q indicates the decomposed matrix A k The value range of the elements in .

[0015] Objective function 1:

[0016]

[0017] Among them, β is the scaling factor, K matrices

[0018] Objective function 2:

[0019]

[0020] Among them, β is the scaling factor, K matrices

[0021] Step 1: Limit A k Sparsity;

[0022] Step 2: Limit A k Element value range;

[0023] Step 3: Limit A k Sparsity and element value range;

[0024] Step 4: Generalization of Kronecker product matrix decomposition;

[0025] Step 5: Limit the error tolerance.

[0026] Preferably, in step 1, the specific method is as follows:

[0027] Step 1.1: N = 2 t , DFT matrix F for t = 1, 2, 3, ... N Under constraint 1, a 0-1 integer programming model is established, and A1, A2, ..., A are solved using a mixed basis. K The optimal solution

[0028] Step 1.2: Establish a 0-1 integer programming model, based on the butterfly idea, and use mixed basis to solve A1, A2, ..., A K The optimal solution of

[0029] Step 1.3: Establish a single-objective optimization model for scaling factor β, and use the butterfly optimization algorithm to solve the optimization schemes for objective functions 1 and 2;

[0030] Step 1.4: Finally, we get the specific optimization design schemes for N=2, 4, 8, and 16. and β, hardware complexity C and error RMSE.

[0031] Preferably, in step 2, the specific method is as follows:

[0032] Step 2.1: For N = 2 t , DFT matrix F for t = 1, 2, 3, 4, 5 N , under constraint 2 and q = 3, restrict the matrix A k The range of real and imaginary parts of the elements in the matrix is ​​used to solve the optimal decomposition matrix A1, A2, ..., A K ;

[0033] Step 2.2: Use the butterfly optimization algorithm to optimize β; finally, we get the specific optimization design schemes for N = 2, 4, 8, 16, 32 and β, hardware complexity C and error RMSE.

[0034] Preferably, in step 3, the specific method is as follows:

[0035] Step 3.1: Consider the matrix A at the same time k The sparsity and value range of the elements, i.e., constraints 1 and 2, are used to establish a 0-1 integer programming model. Based on the butterfly idea, mixed basis is used to solve A1, A2, ..., A K The optimal solution of

[0036] Step 3.2: Establish a single-objective optimization model for β and solve it using the butterfly optimization algorithm; finally, we get the specific optimization solutions for N=2, 4, 8, 16, 32 and β, hardware complexity C and error RMSE.

[0037] Preferably, in step 4, the specific method is as follows:

[0038] Step 4.1: Matrix It is a DFT-like matrix. Based on the generator idea, the honey badger algorithm is used to find the optimal generator α to fit the matrix F N ;

[0039] Step 4.2: According to the butterfly calculation rule, get the integer sparse matrix A1, A2, ..., A K

[0040] Step 3: Honey Badger Algorithm optimizes β: Get a specific optimization solution and β, hardware complexity C and error RMSE.

[0041] Preferably, in step 5, the specific method is as follows:

[0042] Step 5.1: Under the condition that constraints 1 and 2 are satisfied at the same time, design an optimization scheme for q, and use the Honey Badger algorithm to iterate the error RMSE until the error is less than 0.1 to obtain the optimal q;

[0043] Step 5.2: Optimize the generators under different N dimensions to obtain the corresponding optimal generator α, thereby generating F under different dimensions N

[0044] Step 5.3: Use the butterfly calculation method to decompose A1, A2, ..., A K , the model complexity C of N=2, 4, 8, 16, 32 is obtained, and the optimal β is obtained through the honey badger algorithm.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. The present invention uses a two-stage search optimization method and the advantages of mixed basis solution to accurately decompose the DFT matrix, and then uses the butterfly optimization algorithm to achieve an efficient solution process and successfully reduce the hardware complexity.

[0047] 2. The present invention uses a three-stage search optimization method, a combination of generator ideas and the honey badger algorithm to find the optimal parameters more quickly and accurately, and uses the butterfly calculation rule to be well decomposed into a sparse matrix to reduce hardware complexity.

[0048] 3. The method of the present invention is fast and accurate and can effectively reduce hardware complexity. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings, so that those skilled in the art can better understand the advantages and features of the present invention, thereby making a clearer definition of the protection scope of the present invention. The embodiments described in the present invention are only a part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present invention.

[0051] Example:

[0052] A fast approximation method for DFT-like matrices based on integer sparse matrix factorization is proposed. A double-disc mixed basis optimization model is used to optimize matrix decomposition: based on the stage search optimization method, the matrix is ​​first optimized using a mixed basis, and then the real-valued matrix scaling factor β is optimized. Under different constraints, a double-butterfly mixed basis optimization model is established to set N=2 t ,t=1,2,3,......DFT matrix FN is approximately decomposed into the form of integer matrix multiplication to reduce the computational complexity, and then the optimal parameters are found under the condition of generalization to the DEF-like matrix and limiting the accuracy tolerance, which can better reduce both the hardware complexity and the error;

[0053] The present invention mainly considers the following two constraints:

[0054] F N is an N-dimensional DFT matrix in the following form:

[0055]

[0056] Where j is the imaginary unit;

[0057] Consider the following two constraints:

[0058] Constraint 1: Limit Each matrix A k Each row of has at most 2 non-zero elements;

[0059] Constraint 2: Limit Each matrix A k Meet the following requirements:

[0060]

[0061] Among them, A k [l, m] represents the matrix A k The element at row l and column m, A k The range of the real and imaginary parts of the elements in , q indicates the decomposed matrix A k The value range of the elements in .

[0062] Objective function 1:

[0063]

[0064] Among them, β is the scaling factor, K matrices

[0065] Objective function 2:

[0066]

[0067] Among them, β is the scaling factor, K matrices

[0068] The following steps are involved:

[0069] Step 1: Limit A k Sparsity N = 2 t , DFT matrix F for t = 1, 2, 3, ... N , only under the condition of constraint 1, an optimization scheme is designed for objective functions 1 and 2, the hardware complexity is reduced by reducing the number of multipliers, and the minimum error and the hardware complexity of the scheme are calculated.

[0070] The specific method is as follows:

[0071] Step 1.1: Establish a 0-1 integer programming model, based on the butterfly idea, and use mixed basis to solve A1, A2, ..., A K The optimal solution of

[0072] The main goal of this step is to reduce hardware complexity, namely:

[0073] Where C is the hardware complexity, q represents the decomposed matrix A kThe value range of the elements in , L represents the number of complex multiplications (the number of complex multiplications is not counted when multiplying with 0, ±1, ±j); let the kth sparse matrix A k The element in row i and column j is in Represents a 0-1 variable, used to count the number of non-zero elements, according to each matrix A k The assumption that they are all symmetric matrices and constraint 1 give the constraint formula:

[0074]

[0075] According to each matrix A k are all reversible, their inverse matrix complexity is low and each matrix A k are orthogonal or approximately orthogonal, then there is a constraint:

[0076] Among them I N is the N-dimensional identity matrix;

[0077] Therefore, the 0-1 integer programming model is established as follows:

[0078]

[0079] Step 1.2: Establish a single-objective optimization model for scaling factor β and solve it using the butterfly optimization algorithm; finally, obtain specific optimization design solutions for N = 2, 4, 8, and 16 and β, β are 1, 0.5, 0.4767, and 0.3344, respectively.

[0080] Step 1.3: Calculate the corresponding error RMSE as 0, 0.866, 0.737, 0.3755 and the hardware complexity C as 0, 0, 0, 128 respectively.

[0081] Step 1.4: Vertical comparison, low-dimensional DFT matrix F N The hardware complexity is 0. As N increases, the accuracy increases. In contrast, the low-dimensional DFT matrix F N Under the two objective functions, the optimal β values ​​are very similar, the minimum error RMSE of objective function 2 is reduced, and the fitting accuracy is improved.

[0082] Step 2: Limit A k Element value range

[0083] For N = 2 t ,t=1,2,3,4,5 DFT matrix F N , under constraint 2 and q = 3, design an optimization solution to limit A kThe range of real and imaginary parts of the elements in the formula is used to reduce the hardware complexity and calculate the minimum error and solution hardware complexity, where A k The range of the real and imaginary parts of the elements

[0084] The specific method is as follows:

[0085] Step 2.1: First, no longer consider the matrix A k The sparsity of the matrix A is restricted k The range of the real and imaginary parts of the elements in the matrix is ​​solved by using the mixed basis to obtain the optimal decomposition matrix A1, A2, ..., A K ;

[0086] The main objectives of this step are the same as above, namely: First, we no longer consider the matrix A k The sparsity of each matrix A k The assumption is that all matrices are symmetric and each matrix A k If constraint 2 is satisfied, then the constraint formula is:

[0087]

[0088] Among them, A k The range of the real and imaginary parts of the elements in P = {0, ±1, ±2, ±3, ±4} is P = {0, ±1, ±2, ±3, ±4}. k are all reversible, their inverse matrix complexity is low and each matrix A k are orthogonal or approximately orthogonal, then there is a constraint:

[0089] Among them I N is the N-dimensional identity matrix;

[0090] Therefore, the integer programming model of this step is established as follows:

[0091]

[0092]

[0093] By limiting the matrix A k The range of the real and imaginary parts of the elements in the matrix is ​​solved by using the mixed basis to obtain the optimal decomposition matrix A1, A2, ..., A k ;

[0094] Step 2.2: Use the butterfly optimization algorithm to optimize β; finally, we get the specific optimization design schemes for N = 2, 4, 8, 16, 32 and β , β are 1, 0.5001, 0.5944, 0.4196, and 0.4281, respectively;

[0095] Step 2.3: Calculate the corresponding error RMSEs, which are 0, 0.866, 0.7205, 0.2374, and 0.2256, and the hardware complexity C, which are 0, 0, 0, 24, and 78, respectively.

[0096] Step 3: Limit A k Sparsity and element range

[0097] For N = 2 t , DFT matrix F for t = 1, 2, 3, 4, 5 N , based on step 2, consider constraint 1 and limit A k Sparsity and value range, for Optimize the design with β and calculate the minimum error and hardware complexity of the solution. The value range is the same as step 2.

[0098] The specific method is as follows:

[0099] Step 3.1: Consider the matrix A at the same time k The sparsity and value range of elements are studied to study the low complexity design scheme of DFT. A 0-1 integer programming model is established, and A1, A2, ..., A are solved by mixed basis based on the butterfly idea. K The optimal solution of

[0100] The main objectives of this step are the same as above, namely: Let the kth sparse matrix A k The element in row i and column j is in Represents a 0-1 variable, used to count the number of non-zero elements, according to each matrix A k The assumption that they are all symmetric matrices and constraint 1 give the constraint formula:

[0101]

[0102] Each matrix A k If the range of the real and imaginary parts of the elements satisfies constraint 2, then the constraint formula is: Among them, A k The range of the real and imaginary parts of the elements in P = {0, ±1, ±2, ±3, ±4}; according to each matrix A k are all reversible, their inverse matrix complexity is low and each matrix A k are orthogonal or approximately orthogonal, then there is a constraint:

[0103] Among them I N is the N-dimensional identity matrix; also consider the matrix A k The sparsity and value range of elements are considered to study the low-complexity design of DFT. The 0-1 integer programming model is established as follows:

[0104]

[0105] Step 3.2: Establish a single-objective optimization model for β and solve it using the butterfly optimization algorithm; finally, we get the specific optimization solutions for N=2, 4, 8, 16, 32 and β , β are 1, 0.49989, 0.6209, 0.5, and 0.5095, respectively;

[0106] Step 3.3: Calculate the corresponding error RMSE as 0, 0.866, 0.719, 0.162, 0.11 and the hardware complexity C as 0, 0, 0, 24, 78 respectively.

[0107] Step 4: Generalization of Kronecker product matrix decomposition

[0108] Based on the above problem, consider the matrix where F N 1 and F N2 are the DFT matrices of dimensions N1 and N2 respectively, represents the Kronecker product. When N1=4,

[0109] When N2=8, please satisfy constraints 1 and 2 at the same time. and β are optimized, and the minimum error and the hardware complexity C of the solution are calculated. The value range is the same as step 2.

[0110] The specific method is as follows:

[0111] Step 4.1: Matrix It is a DFT-like matrix. Based on the generator idea, the honey badger algorithm is used to find the optimal generator α to fit the matrix F N ;

[0112] The DFT-like matrix F N1 and F N2 The DFT matrices are N1=4 and N2=8 respectively. Since they are multiples of 4, they have strong symmetry properties, so the following relationship holds:

[0113]

[0114] The matrix mapping for this question is defined as follows:

[0115] The generator α=(a0,a2,…,a N / 4-1 ) T express dimensional complex column vector, is the fitting matrix,

[0116] The relationship between the internal elements of the matrix and the generators is:

[0117] This step is based on the generator idea, and the honey badger algorithm is used to find the optimal generator α to fit the matrix F N ;

[0118] Define the fitting matrix With the DFT matrix F N The distance is:

[0119]

[0120] Therefore, in the process of minimizing d(α), the Honey Badger algorithm can find an optimal generator α to fit the target matrix The complex vector α satisfies the condition: a n ∈P 2 , where A k The range of the real and imaginary parts of the elements in is P = {0, ±1, ±2, ..., ±2q-1}; and a n can be expressed in the form of: Therefore, the optimization objective function of this step is:

[0121] Step 4.2: According to the butterfly calculation rule, get the integer sparse matrix A1, A2, ..., A K , for sparse matrices Decompose, as above, set the kth sparse matrix A k The element in row i and column j is in Represents a 0-1 variable, used to count the number of non-zero elements, according to each matrix A k Assumption and constraint 1 that all matrices are symmetric; each matrix A k The range of the real and imaginary parts of the elements satisfies constraint 2; according to each matrix A k are all reversible, their inverse matrix complexity is low and each matrix A k If they are all orthogonal or approximately orthogonal, then the constraint formula is:

[0122]

[0123]

[0124] Fitting Matrix It can be defined as: And satisfy the fitting matrix The inverse matrix of is of low complexity, and the fitting matrix Should satisfy orthogonality or nearly orthogonality;

[0125] Step 4.3: The result obtained in step 2 is recorded as The honey badger algorithm optimizes β, and the fitness function is:

[0126]

[0127] Step 4.4: Get the specific design scheme: the optimal real-valued matrix scaling factor β is 0.4093, the minimum error is 0.2177, and the hardware complexity is 48.

[0128] Step 5: Limit error tolerance

[0129] Based on the above problems, considering the accuracy requirement error is less than 0.1, for N = 2 t , t=1,2,3... DFT matrix F N , please satisfy constraints 1 and 2 at the same time. and Perform optimization and calculate the hardware complexity C of the solution.

[0130] The specific method is as follows:

[0131] Step 5.1: Design an optimization scheme for q, and use the RMSE error of the Honey Badger algorithm to iteratively find the optimal q value and search The generator of q refers to the decomposed matrix A k The range of elements in A k About, A k It is related to RMSE, that is, q is related to RMSE, and finally the optimal q=6 is obtained;

[0132] Step 5.2: Optimize the generators under different N dimensions to obtain the corresponding optimal generator α, thereby generating F under different dimensions N , define the fitting matrix With the DFT matrix F N The distance is: Therefore, in the process of minimizing d(α), the Honey Badger algorithm can find an optimal generator α to fit the target matrix The complex vector α satisfies the following conditions: n ∈P 2 , where A k The range of the real and imaginary parts of the elements in is P = {0, ±1, ±2, ..., ±2q-1}; and a n can be expressed in the form of: Therefore, the optimization objective function of this step is:

[0133] Step 5.3: Use the butterfly calculation method to decompose A1, A2, ..., A KAccording to the butterfly calculation rule, we get the integer sparse matrix A1, A2, ..., A K , for sparse matrices Decompose, as above, set the kth sparse matrix A k The element in row i and column j is in Represents a 0-1 variable, used to count the number of non-zero elements, according to each matrix A k Assumption and constraint 1 that all matrices are symmetric; each matrix A k The range of the real and imaginary parts of the elements satisfies constraint 2; according to each matrix A k are all reversible, their inverse matrix complexity is low and each matrix A k If they are all orthogonal or approximately orthogonal, then the constraint formula is:

[0134]

[0135]

[0136] Defined as: And satisfy the fitting matrix The inverse matrix of is of low complexity, and the fitting matrix Orthogonality or approximate orthogonality should be satisfied; the model complexity C of N=2, 4, 8, 16, 32 is 0, 0, 0, 48, 156 respectively;

[0137] Step 5.4: The optimal β values ​​obtained by the honey badger algorithm are 1, 0.5, 0.348, 0.2488, and 0.1757.

[0138] In summary, the present invention establishes a double-butterfly hybrid basis optimization model under different constraints. t , DFT matrix F for t = 1, 2, 3, ... N Approximate decomposition into integer matrices A1, A2, ..., A K The form of continuous multiplication is used to reduce the computational complexity, and then the optimal parameters are found when it is extended to the DEF-like matrix and the accuracy tolerance is limited, which can better reduce both the hardware complexity and the error.

[0139] The description and practice disclosed in the present invention are easy to think and understand for ordinary technicians in the technical field, and several improvements and modifications can be made without departing from the principles of the present invention. Therefore, modifications or improvements made without departing from the spirit of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A fast approximation method for DFT-like matrices based on integer sparse matrix factorization, using a double-disc mixed basis optimization model to optimize matrix decomposition: Based on the stage search optimization method, the matrix is ​​first optimized using a mixed basis, and then the real-valued matrix scaling factor β is optimized. Under different constraints, a double-butterfly mixed basis optimization model is established to set N=2 t , DFT matrix F for t = 1, 2, 3, ... N Approximately decompose into the form of integer matrix multiplication to reduce the computational complexity, and then find the optimal parameters when generalizing to the DEF-like matrix and limiting the accuracy tolerance, which is characterized by: The following steps are involved: F N is an N-dimensional DFT matrix in the following form: Where j is the imaginary unit; Consider the following two constraints: Constraint 1: Limit Each matrix A k Each row of has at most 2 non-zero elements; Constraint 2: Limit Each matrix A k Meet the following requirements: Among them, A k [l, m] represents the matrix A k The element at row l and column m, A k The range of the real and imaginary parts of the elements in , q indicates the decomposed matrix A k The value range of the elements in; Objective function 1: Among them, β is the scaling factor, K matrices Objective function 2: Among them, β is the scaling factor, K matrices Step 1: Limit A k Sparsity; Step 2: Limit A k Element value range; Step 3: Limit A k Sparsity and element value range; Step 4: Generalization of Kronecker product matrix decomposition; Step 5: Limit the error tolerance.

2. The method for fast approximation of DFT-like matrices based on integer sparse matrix factorization according to claim 1, characterized in that: In step 1, the specific method is as follows: Step 1.1: N = 2 t , DFT matrix F for t = 1, 2, 3, ... N Under constraint 1, a 0-1 integer programming model is established, and A1, A2, ..., A are solved using a mixed basis. K The optimal solution of Step 1.2: Establish a 0-1 integer programming model, based on the butterfly idea, and use mixed basis to solve A1, A2, ..., A K The optimal solution of Step 1.3: Establish a single-objective optimization model for scaling factor β, and use the butterfly optimization algorithm to solve the optimization schemes for objective functions 1 and 2; Step 1.4: Finally, we get the specific optimization design schemes for N=2, 4, 8, and 16. and β, hardware complexity C and error RMSE.

3. The method for fast approximation of DFT-like matrices based on integer sparse matrix factorization according to claim 1, characterized in that: In step 2, the specific method is as follows: Step 2.1: For N = 2 t , DFT matrix F for t = 1, 2, 3, 4, 5 N , under constraint 2 and q = 3, restrict the matrix A k The range of the real and imaginary parts of the elements in the matrix is ​​solved by using the mixed basis to obtain the optimal decomposition matrix A1, A2, ..., A K ; Step 2.2: Use the butterfly optimization algorithm to optimize β; finally, we get the specific optimization design schemes for N=2, 4, 8, 16, 32. and β, hardware complexity C and error RMSE.

4. The method for fast approximation of DFT-like matrices based on integer sparse matrix factorization according to claim 1, characterized in that: In step 3, the specific method is as follows: Step 3.1: Consider the matrix A at the same time k The sparsity and value range of the elements, i.e., constraints 1 and 2, are used to establish a 0-1 integer programming model. Based on the butterfly idea, mixed basis is used to solve A1, A2, ..., A K The optimal solution of Step 3.2: Establish a single-objective optimization model for β and solve it using the butterfly optimization algorithm; finally, we get the specific optimization solutions for N=2, 4, 8, 16, 32 and β, hardware complexity C and error RMSE.

5. The method for fast approximation of DFT-like matrices based on integer sparse matrix factorization according to claim 1, characterized in that: In step 4, the specific method is as follows: Step 4.1: Matrix It is a DFT-like matrix. Based on the generator idea, the honey badger algorithm is used to find the optimal generator α to fit the matrix F N ; Step 4.2: According to the butterfly calculation rule, get the integer sparse matrix A1, A2, ..., A K ; Step 4.3: Honey Badger Algorithm optimizes β: Get a specific optimization solution and β, hardware complexity C and error RMSE.

6. The method for fast approximation of DFT-like matrices based on integer sparse matrix factorization according to claim 1, characterized in that: In step 5, the specific method is as follows: Step 5.1: Under the condition that constraints 1 and 2 are satisfied at the same time, design an optimization scheme for q, and use the Honey Badger algorithm to iterate the error RMSE until the error is less than 0.1 and the optimal solution is obtained; Step 5.2: Optimize the generators under different N dimensions to obtain the corresponding optimal generator α, thereby generating F under different dimensions N ; Step 5.3: Use the butterfly calculation method to decompose A1, A2, ..., A K , the model complexity C of N=2, 4, 8, 16, 32 is obtained, and the optimal β is obtained through the honey badger algorithm.

Citation Information

Cited By

  • Method and device for establishing feature factorization remote sensing target counting model

    CN120912555A