A Method and System for Generating Low Peak-to-Average Power Ratio Transmit Waveforms for AFDM Based on Precise Search of Candidate Upper and Lower Bounds
By using candidate upper and lower bounds and a global scheduling method in AFDM signal processing, the problem of high computational cost in strict exhaustive CSM is solved, achieving efficient AFDM transmit waveform generation and reducing computational complexity and resource consumption.
Patent Information
- Application Number
- CN202611249973.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-18
- Publication Date
- 2026-11-17
AI Technical Summary
In existing AFDM signal processing techniques, the strict exhaustive use of the CSM method to generate AFDM transmit waveforms results in a large computational load and an inability to effectively utilize the samples generated during the candidate selection process, leading to redundant calculations and wasted resources.
By maintaining valid lower and upper bounds for candidates within the upper and lower bounds, global scheduling is performed using candidate numbers and sample access order to determine winning candidates in advance, and precise advancement is continued when necessary. The inverse transformation calculation state is reused to avoid repeated calculation of already generated samples and butterfly nodes.
This reduces computational complexity and resource consumption, ensures that the generated AFDM transmit waveform is consistent with strict exhaustive search, and improves computational efficiency and resource utilization.
Smart Images

Figure CN123123090A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of wireless communication baseband signal processing, Affine Frequency Division Multiplexing (AFDM) transmit waveform generation, and Peak-to-Average Power Ratio (PAPR) control, specifically to chirp-parameter selected mapping (CSM) in AFDM transmitters. More specifically, this invention relates to a method and system for accurately searching and generating complete low PAPR transmit waveforms based on upper and lower bounds of a finite AFDM-CSM candidate set. Background Technology
[0002] AFDM generates the transmit signal using Discrete Affine Fourier Transform (DAC) and can be used for information transmission in time-frequency dual-selective channels. Similar to other multicarrier waveforms, the multiple complex components of the AFDM transmit signal, when superimposed in the time domain, may exhibit a high instantaneous peak power (PAPR). A higher PAPR increases the linear operating range requirements of subsequent transmit chains and may necessitate a larger operating margin in the power amplification stage.
[0003] CSM sets a set of discrete second chirp parameters for the same input data symbol block, forming multiple equal-energy AFDM candidates, and selects the candidate with the smallest PAPR on a specified discrete oversampled target grid. Each candidate carries the same data, and their differences come from the unit modulus quadratic phase specified in the candidate codebook. Therefore, the purpose of candidate selection is not to change the input information, but to determine the transmit waveform with the smaller peak value of the target grid from a finite set of candidates.
[0004] Let the length of the AFDM block be... The number of candidates is Oversampling factor is Strictly exhaustively list all CSMs that need to be generated and evaluated. The process involves generating candidate target mesh composite points and then comparing the complete target mesh PAPR of each candidate. While a rigorous exhaustive search yields a definite finite mesh selection result, before obtaining the final winning candidate, a subset of candidate sample points that do not actually contribute to the final winning relationship are generated, and the inverse transform butterfly nodes associated with these sample points are calculated.
[0005] While reducing the computational load, the candidate set, target grid, PAPR definition, and stalemate rules must remain unchanged. If candidate competition is advanced only based on a subset of sample points, the sample points used should be exact composite samples on the target grid; the lower bound of the candidate should not be higher than the true target value, and the upper bound of the candidate should not be lower than the true target value. Furthermore, the optimality criteria should be verifiable, consisting of valid upper and lower bounds and their comparison with all competing candidates. For ease of description, these optimality criteria will also be referred to as the winning certificate below. When the criteria cannot be formed, further precise calculations are required, and unproven approximate candidates should not be used as the final result.
[0006] When a winning candidate is determined in advance, all candidates may not yet have been generated. The target mesh is a complex sample point. If the winning candidate is reconstructed and the complete inverse transform is performed from scratch to obtain the complete transmitted waveform, the sample points generated in the selection phase and the butterfly nodes already calculated will not be reused, and the work saved in the candidate selection phase will be reintroduced in the waveform formation phase. Therefore, the selection process and the winning candidate waveform completion process need to be connected into the same pauseable and resumable computational chain. Summary of the Invention
[0007] The purpose of this invention is to provide a method and system for generating AFDM low peak-to-average power ratio transmit waveforms based on precise search of candidate upper and lower bounds, in order to solve the following technical problems: under the condition of not changing the candidate set, fixed target grid, objective function and minimum candidate number priority tie rule of strict exhaustive search CSM, how to maintain effective lower and upper bounds based on candidate target grid samples that have not yet been fully generated, determine the result in advance by covering all competing candidates and conforming to the tie direction, and continue to precisely advance until finite termination when the certificate cannot be formed, thereby ensuring that the obtained winning candidate is consistent with strict exhaustive search.
[0008] Building upon this foundation, the present invention also addresses: how to generate precise replica points on a fixed target grid according to a global scheduling request and reuse already completed inverse transform calculations; and how, when the winning candidate is not yet complete, to continue reusing the inverse transform calculation state within the same candidate retained at the time of stopping selection, generating only missing samples, avoiding recalculating the winning candidate from scratch, and ultimately outputting all samples. The complete inner layer complex waveform of each sample point.
[0009] To achieve the above objectives, the present invention adopts the following technical solution:
[0010] A method for generating low peak-to-average power ratio transmit waveforms in simulated radio frequency demultiplexing (AFDM) based on precise candidate upper and lower bound search includes the following steps:
[0011] (a) Obtain the complex data symbol block to be sent, construct a finite number of AFDM chirp selection mapping candidates for the data symbol block, define the inner layer complex sampling point and the peak-to-average power ratio of the target grid for each candidate on a fixed target grid, and stipulate that when the peak-to-average power ratio of the target grid is equal, the candidate with the smaller number is selected.
[0012] (b) For each candidate, maintain a candidate lower bound that is not higher than the peak-to-average power ratio of the candidate's true target grid and a candidate upper bound that is not lower than the peak-to-average power ratio of the true target grid, set the initial mathematical values of the candidate lower bound and the candidate upper bound, and specify the target grid sample access order for all candidates to use.
[0013] (c) Establish candidate inverse transformation calculation states that can be paused and resumed for each candidate, generate accurate composite sample points of the target grid as needed according to the target grid sample point access order, and reuse the completed inverse transformation calculations;
[0014] (d) Schedule the candidates according to the dictionary key consisting of the candidate lower bound and the candidate number, generate a target grid exact sample point that has not been visited before, and update the candidate lower bound and candidate upper bound of the selected candidate according to the generated sample points;
[0015] (e) Based on the comparison between the upper bound of each candidate and the lower bound of all other candidates, determine whether the winning candidate can be determined; if not, return to step (d) to continue generating sample points and updating the upper and lower bounds until the winning candidate is determined.
[0016] (f) After determining the winning candidate, stop advancing other candidates, retain the candidate inverse transformation calculation state of the winning candidate, and continue to generate target mesh exact replica points that have not yet been visited only from the retained state, without reconstructing the winning candidate, without regenerating existing sample points, and without recalculating cached nodes;
[0017] (g) After the winning candidate generates all the target mesh exact replica points, they are arranged in natural time index order to form a complete inner layer complex waveform.
[0018] Furthermore, in step (a), the length of the data symbol block... and oversampling factor The determined length is On a fixed target grid, the peak-to-average power ratio of each candidate inner-layer complex sample point and the target grid is defined, and the complex data symbol block is represented as... ,in, , Set the oversampling factor , and The number of candidates is a positive integer. Not less than An integer, given a finite real base second chirp parameter And for candidate numbers ,according to
[0019]
[0020] structure AFDM chirp selection mapping candidates; let
[0021]
[0022] in, and the first A candidate inner layer replica point on the fixed target mesh is defined as...
[0023]
[0024] by
[0025]
[0026] As the first Peak-to-average power ratio of the target grid for each candidate.
[0027] Furthermore, in step (b), the lower bound of each candidate... and candidate upper bound satisfy The initial mathematical value of the candidate lower bound is The initial mathematical value of the candidate upper bound is positive infinity; the target grid sample point access order is as follows:
[0028]
[0029] in, Indicates to The bitwise binary representation performs bit reversal, and the access sequence forms a complete uniform subgrid at a preset milestone. ,satisfy
[0030]
[0031] So that the access order is before Each position constitutes a complete Double uniform subgrid.
[0032] Furthermore, in step (c), the natural time index of the target grid is decomposed into... ,in, , Make each remainder class Corresponding to a length of The inverse discrete Fourier transform; only when the first request belongs to the remainder class. When the target mesh is accurately replicated, the inert radix-2 inverse transform graph of the remainder class is initiated, and a sequence is generated.
[0033]
[0034] The length of this remainder class is The inverse discrete Fourier transform input is used; when subsequent requests are made for exact sample points of the target mesh belonging to the same remainder class, the same inert radix-2 inverse transform graph is restored and the butterfly nodes already calculated therein are reused; the candidate intra-inverse transform calculation state includes the current position in the target mesh sample point access order, the identifier of the activated remainder class, the calculated butterfly nodes of each activated remainder class, and the natural time index. With the corresponding remainder class Remainder class index The mapping relationship between the terminal nodes of the inert radix-2 inverse transform graph, the generated labels of the target grid's accurate complex points, and the currently observed maximum sample power.
[0035] Furthermore, in step (d), after the selected candidates have been generated... After accurately replicating each target grid point, according to
[0036]
[0037] Update the lower bound of the selected candidates.
[0038] Furthermore, in step (d), when the number of generated sample points reaches a non-final preset milestone, the candidate upper bound of the candidate is updated using the precise complex sample points on the corresponding complete uniform subgrid; when the number of generated sample points reaches the final preset milestone... When the candidate's lower bound and upper bound are updated to their true target grid peak-to-average power ratio;
[0039] Wherein, when the number of samples generated by the selected candidate according to the target grid sample point access order reaches the complete Preset milestones corresponding to double uniform subgrids season
[0040]
[0041] And in Time calculation
[0042]
[0043] exist season The candidate upper bound of the selected candidate is updated with the smaller of the currently saved candidate upper bound and the candidate upper bound calculated in this operation.
[0044] Furthermore, in step (e), any candidate with a finite upper bound is taken as a candidate to be judged; when the upper bound of the candidate to be judged is strictly less than the lower bound of all other candidates whose numbers are less than the candidate to be judged, and is less than or equal to the lower bound of all other candidates whose numbers are greater than the candidate to be judged, the candidate to be judged is determined as the winning candidate.
[0045] Among them, for candidates to be determined with a finite upper bound When both conditions are met
[0046]
[0047] At that time, candidates The candidate is determined as the winning candidate; the finite number of candidates is denoted as... Each time a candidate is advanced but not yet completed, only one exact replica point of the target mesh that has not been visited before is generated. Each candidate generates at most one exact replica point of the target mesh. The aforementioned sample points, and at the latest generated by accumulating all candidate samples. The winning candidate is determined when the number of the aforementioned sample points is reached.
[0048] Furthermore, in steps (f) and (g), let This represents the number of exact replica points of the target mesh that the winning candidate has generated when candidate selection stops, where, ;when At that time, the inverse transformation within the candidate domains of the retained winning candidates is used to calculate the state, generating only those that have not yet been visited. Each target grid is accurately replicated at multiple points; when At that time, no new target mesh exact replica points or new radix-2 butterfly nodes will be generated; in all of the winning candidates After generating accurate replica points for each target grid, according to The order is determined by and Determine the cached terminal nodes in the lazy radix-2 inverse transform graph of the corresponding remainder class, read the terminal nodes and follow the... Normalization is performed to assemble the complete inner complex waveform; the reading, normalization, and assembly do not trigger new radix-2 butterfly node calculations, butterfly complex multiplications, or butterfly complex additions.
[0049] An AFDM low peak-to-average power ratio transmit waveform generation system based on precise candidate upper and lower bound search includes:
[0050] The candidate construction module is used to construct a finite number of AFDM chirp selection mapping candidates. It defines the inner layer complex points of each candidate and the peak-to-average power ratio of the target grid on a fixed target grid, and stipulates that when the peak-to-average power ratios of the target grids are equal, the candidate with the smaller number is selected.
[0051] The upper and lower bound maintenance module is used to maintain the candidate lower bound and candidate upper bound for each candidate, set the initial mathematical values of the candidate lower bound and candidate upper bound, and specify the target grid sample point access order for all candidates to use.
[0052] The inertial inverse transformation calculation module is used to establish candidate intra-candidate inverse transformation calculation states that can be paused and resumed for each candidate, generate accurate composite sample points of the target grid as needed according to the target grid sample point access order, and reuse the completed inverse transformation calculations.
[0053] The scheduling and lower bound update module is used to schedule candidates according to the dictionary key consisting of the candidate lower bound and the candidate number, generate a target grid exact resample point that has not been visited before, and update the candidate lower bound of the selected candidate based on the generated sample points.
[0054] The upper bound update module is used to update the upper bound of the selected candidates based on the generated sample points.
[0055] The optimality determination module is used to determine the winning candidate based on the comparison relationship between the upper bound of each candidate and the lower bound of all other candidates. If it cannot be determined, it triggers the generation of more samples and updates the upper and lower bounds until the winning candidate is determined.
[0056] The state reuse and completion module is used to retain the inverse transform calculation state within the candidate after the winning candidate is determined, and to continue generating the target mesh exact replica points that have not yet been visited only from the retained state;
[0057] The waveform assembly output module is used to arrange the complete inner layer complex waveform in natural time index order after the winning candidate generates all the target grid accurate replica points.
[0058] Beneficial effects:
[0059] Compared with the prior art, the present invention has the following beneficial effects.
[0060] (1) This method describes candidates that have not yet been fully evaluated with always valid candidate lower bounds and candidate upper bounds, and forms a verifiable winning certificate by finite upper bounds, all competing candidate lower bounds and corresponding comparison relationships, without replacing the selected target on the finite target grid with coarse grid approximation.
[0061] (2) The candidate lower bound and the candidate number jointly participate in the global scheduling, so that the computation is given priority to the candidate with the smaller current dictionary key; the common bit inversion prefix not only specifies the order of accessing the precise sample points, but also provides the complete uniform subgrid required to update the candidate upper bound at the milestone.
[0062] (3) The strict and non-strict comparison direction in the winning certificate is consistent with the tie rule of minimum candidate number priority; when a certificate cannot be formed in advance, the process continues to proceed precisely and terminates on a finite target grid, thus maintaining the same winning candidate and tie result as strict exhaustive search.
[0063] (4) The candidate inert radix-2 inverse transformation state serves as the sample calculation carrier for precise search, and can be paused and resumed multiple times. Subsequent sample requests can reuse the already calculated butterfly nodes, avoiding repeated calculations on the common dependent nodes of the same candidate.
[0064] (5) After stopping candidate selection, only the winning candidate remains in the same state, without regenerating existing samples or recalculating cached butterfly nodes. The number of samples to be completed is determined as follows: After completion, a complete inner-layer complex waveform is formed according to the natural time index, which can be used by the subsequent AFDM transmit chain. Attached Figure Description
[0065] Figure 1 This invention provides an overall flowchart for maintaining upper and lower bounds of AFDM-CSM candidates, performing global scheduling and generating winning certificates, and then reusing the winning candidate states to fill in missing samples and output complete inner complex waveforms.
[0066] Figure 2 Under a fixed experimental configuration, when strictly exhaustively enumerating the CSM and using this method to form a complete inner complex waveform, the average count per data block is compared for the number of complex multiplications, the number of complex additions, and the number of target grid complex sample points generated. Each item is calculated with the average count of the strictly exhaustive enumeration as 100%. The total count for each item in this method includes two consecutive stages: candidate selection and completion by reusing the winning candidate state. Detailed Implementation
[0067] The invention will now be further explained with reference to the accompanying drawings.
[0068] This invention discloses a method for generating low peak-to-average power ratio (PAPS) transmit waveforms in simulated radio frequency multiplexing (AFDM) based on precise candidate upper and lower bound search. A finite number of CSM candidates are constructed for the same data symbol block, and within the same length... PAPR is evaluated on a discrete oversampled target grid, and when PAPR is the same, the candidate with the smaller number is selected. The AFDM block length and oversampling factor are respectively set to... and To support the uniform subgrid prefix property of reversing the access order of common bits, and in the current implementation, the target grid complex points are generated using a radix-2 inverse transform graph organized by remainder class.
[0069] For each candidate, a lower bound is maintained that is no higher than the true PAPR, and an upper bound is maintained that is no lower than the true PAPR. The initial mathematical values for the upper bound are 1 and positive infinity, respectively. All candidates are accessed in reverse order using the same common bit. The lower bound of the candidates is updated based on the currently observed maximum sample power, and candidates are globally scheduled according to a lexicographical key composed of the lower bound and the candidate number. The common access order forms a complete uniform subgrid at a specified milestone. When a non-final preset milestone is reached, the maximum normalized sample power on that complete subgrid is used to form a valid upper bound of the candidates, and the saved upper bound is updated to the smaller of the current value and the current valid upper bound.
[0070] For any candidate with a finite upper bound, it is considered a candidate to be judged. Other candidates with numbers lower than the candidate to be judged are treated as strictly less than, and other candidates with numbers higher than the candidate to be judged are treated as less than or equal to. When the upper bound of the candidate to be judged and the lower bounds of all other candidates simultaneously satisfy the corresponding comparison relationships, the candidate to be judged is determined as the winning candidate. When any candidate generates all... When there are 1000 precise sample points, both the lower and upper bounds of the candidate are set to the peak-to-average power ratio of the actual target grid, and they participate in the winning candidate determination according to the same comparison relationship.
[0071] If no winning certificate has been obtained, a candidate is selected again by lexicographical key, and an unvisited exact replica point of that candidate is requested. Each non-terminating advance adds one generated sample point to one of the finite candidates; each candidate generates at most one generated sample point. Each sample point, therefore the selection process must be completed at least by the time the data is generated. The process ends when all sample points are reached. The resulting winning candidate and tie results, prioritizing the candidate with the smaller number, are consistent with strict exhaustive search.
[0072] To execute the precise replica point request, the natural time index of the target grid is decomposed into remainder classes and indices within remainder classes. For each candidate, a lazy radix-2 inverse transform state is established, capable of pausing, resuming, and saving already computed nodes. When a sample point of a remainder class is requested for the first time, the corresponding inverse transform graph is initiated; when the same remainder class is requested again, the graph is restored, existing valid nodes are reused, and only nodes necessary for generating the target terminal node that are not yet valid are computed. Thus, each candidate advancement generates a previously unvisited precise target grid replica point, and the inverse transform state serves as the computational carrier for the samples required for the aforementioned upper and lower bound updates, global scheduling, and the formation of the winning certificate.
[0073] Once the winning candidate is determined, proceeding to other candidates is stopped, and the same inverse transform state as the winning candidate is retained. Let the winning candidate already exist at the time of stopping. A precise replication point; when Only complete at this time A missing sample point, when The completion is skipped. After all samples are generated, the cached terminal nodes are read and normalized according to the natural time index to form the complete inner complex waveform. The completion does not reconstruct candidates, and the final reading and assembly do not re-execute the already completed inverse transform calculation.
[0074] In some embodiments, the present invention provides a method for generating low peak-to-average power ratio transmit waveforms in simulated radio frequency demultiplexing (AFDM) based on precise candidate upper and lower bound search, comprising the following steps:
[0075] (a) Obtain the complex data symbol block to be sent, construct a finite number of AFDM chirp selection mapping candidates for the data symbol block, and determine the length of the data symbol block. and oversampling factor The determined length is On a fixed target grid, define the inner layer composite points of each candidate and the peak-to-average power ratio of the target grid, and stipulate that when the peak-to-average power ratios of the target grids are equal, select the candidate with the smaller number.
[0076] (b) For each candidate, maintain a candidate lower bound that is not higher than the peak-to-average power ratio of the candidate real target grid and a candidate upper bound that is not lower than the peak-to-average power ratio of the real target grid. Set the initial mathematical values of the candidate lower bound and the candidate upper bound, and specify the target grid sample access order for all candidates to use. The access order forms a complete uniform subgrid at a preset milestone.
[0077] (c) Establish candidate inverse transformation calculation states that can be paused and resumed for each candidate, generate target grid exact complex samples for candidate exact search as needed according to the target grid sample access order, and reuse the inverse transformation calculations already completed in the candidate inverse transformation calculation states.
[0078] (d) Select the candidate with the smallest key from all candidates according to the lexicographical key consisting of the lower bound of the candidate and the candidate number; if the selected candidate has not yet generated all candidates... When generating a target mesh accurate replica point, another previously unvisited target mesh accurate replica point is generated from the candidate's inverse transform calculation state. The candidate's lower bound is updated based on the already generated target mesh accurate replica points. When the number of generated sample points reaches a non-final preset milestone, the candidate's upper bound is updated using accurate replica points on the corresponding complete uniform subgrid. When the number of generated sample points reaches the final preset milestone... When the candidate's lower bound and upper bound are updated to their true target grid peak-to-average power ratio;
[0079] (e) For any candidate with a finite upper bound, take it as a candidate to be judged; when the upper bound of the candidate to be judged is strictly less than the lower bound of all other candidates whose numbers are less than the candidate to be judged, and is less than or equal to the lower bound of all other candidates whose numbers are greater than the candidate to be judged, the candidate to be judged is determined as the winning candidate; if the winning candidate cannot be determined yet, return to step (d) to continue generating accurate replica points of the target grid and updating the upper and lower bounds of the candidates until a winning candidate consistent with strict exhaustive search is obtained within the finite target grid;
[0080] (f) After determining the winning candidate, stop advancing other candidates, save the selection results, and retain the inverse transformation calculation state within the candidate when the candidate selection was stopped; if the winning candidate has not yet been fully generated... When generating a target mesh exact replica point, only the target mesh exact replica point that has not yet been visited is generated from the same candidate inverse transform calculation state that has been retained. The winning candidate is not reconstructed, the target mesh exact replica point that has already been generated is not requested again, and the cached inverse transform node is not recalculated.
[0081] (g) When all winning candidates are generated After accurately replicating each target grid point, the target grid points are arranged in natural time index order to form a complete inner complex waveform that serves as the input to the AFDM transmit chain.
[0082] The following is combined Figure 1 , Figure 2 The invention will be further illustrated with specific embodiments. The single experimental configuration described is used to illustrate feasibility and technical effects, and the parameter values and cost reductions therein are not limited to other implementation scenarios. The candidate selection, winning state completion, and waveform assembly described below are consecutive processing procedures within the same data symbol block.
[0083] Example
[0084] I. Candidate and Target Grids and Target Selection
[0085] Let the length of the AFDM block be... The target grid oversampling factor is ,in and It is a positive integer; the number of candidates is a finite integer. Candidate number is .make For length is Complex data symbol blocks, For a given finite real reference second chirp parameter, the... The second chirp parameter is
[0086]
[0087] Formed according to the second chirp parameter of the benchmark. and form the first 1 candidate frequency domain input sequence :
[0088]
[0089] in . No. The candidates are in length of The resampled points on the inner layer oversampled target mesh are
[0090]
[0091] All candidates apply a second-order phase of unit modulus. Combined with the orthogonality of the target mesh, all candidates share a common average energy.
[0092]
[0093] No. The linear PAPR of each candidate and the winning candidate number are as follows:
[0094]
[0095] It can be known from the fact that the maximum sample power is not less than the average sample power of the target grid Therefore, setting the initial mathematical value of the candidate lower bound to 1 will not exceed the true target value of any candidate. Since all candidates use a common... The differences in the candidate competition come from the maximum sample power on their respective target grids, rather than from the different energy normalization calibers.
[0096] The term "precise selection" as used in this invention refers to the selection within the aforementioned limited range. On the discrete target grid, the same winning candidate is obtained as with strict exhaustive search; the same occurrence Choosing the smallest number at any given time does not guarantee a globally optimal decision for PAPR over continuous time.
[0097] II. Candidate Upper and Lower Bounds, Global Scheduling, and Precise Search Framework
[0098] For each candidate, maintain the following:
[0099]
[0100] candidate lower bound and candidate upper bound The initial values for both are 1 and positive infinity, respectively. Candidate It has been generated in the public access order described in Section 3. When there are a precise number of replicate points, its lower bound is updated to
[0101]
[0102] The observed sample set is a subset of the complete target grid sample set; therefore, the maximum normalized sample power of the observed samples cannot exceed the maximum normalized sample power of the complete target grid. With... As the maximum value increases, it can only remain the same or increase further. Therefore, this candidate lower bound is monotonically non-decreasing and never exceeds the true value. .
[0103] From all candidates, sorted by lexicographical key The candidate with the smallest key is selected; when the lower bounds are the same, the one with the smaller number is preferred. This scheduling relationship ensures that the object with the smaller current lower bound, which is still more likely to become the smallest PAPR candidate under the existing information, gets the next precise sample first, while maintaining the tie-breaking direction of the smallest number from the scheduling stage to the final winner determination.
[0104] If not all of the selected candidates have been generated If a target grid is precisely replicated, the next unvisited precisely replicated point in the common access order is requested, and the sample point is generated according to the candidate recoverable inverse transform state described in Section 5. Then, the upper and lower bounds of the candidate are updated according to the milestone rules described in Section 3. After each update, the winning candidate is determined according to the unified comparison relationship described in Section 4.
[0105] III. Reversing the access order of common bits and determining the upper bound of candidates
[0106] All candidates share the same target grid access order.
[0107]
[0108] in Indicates to Bitwise binary representation performs bit reversal. For The order satisfies
[0109]
[0110] Formerly Each access location constitutes a complete Double-uniform subgrid. Bit inversion only specifies the sample generation order and does not change the sample values, candidate set, or PAPR target. For this embodiment... and The complete subgrid milestones appear sequentially when 128, 256, 512, 1024, and 2048 sample points have been generated.
[0111] Before candidate completion When calculating individual sample points...
[0112]
[0113] No. After continuous interpolation of each candidate frequency is centered using a unit modulus phase without changing the amplitude, its frequency index is located at... Within this framework, the adopted band-limited peak-to-peak relationship is satisfied; simultaneously, the target grid sample access sequence forms a complete sequence at each preset milestone. Double the uniform subgrid, and Therefore, in At each non-final preset milestone, an upper bound for the effective candidates is formed.
[0114]
[0115] This upper bound is based on the peak values on a fully uniform subgrid and utilizes the band-limited peak-to-peak relationship after frequency centering to cover the peak values of the entire target grid. Its purpose is to provide a value that does not underestimate the true value. Instead of directly replacing the full target mesh PAPR with the peak values of the sparser subgrids, the finite values are used.
[0116] when season The saved candidate upper bound is the smaller of the current value and the currently valid upper bound; although the upper bound coefficient varies with... As the grid becomes larger and tighter, the measured peak value of the subgrid may still increase. Therefore, the smaller value of the effective upper bound of each previous test is used to maintain the tightest effective result that has been obtained.
[0117] IV. Optimality Criteria and Finite Termination Consistent with a Draw
[0118] This invention employs a unified comparison relationship between candidate upper and lower bounds to form the winning certificate. Any candidate with a finite and valid upper bound can be used as a candidate for judgment; when a candidate reaches the final milestone, its candidate upper and lower bounds are equal to the peak-to-average power ratio of the actual target grid according to the rules described in Sections 2 and 3, and it continues to participate in the optimality determination according to the same comparison relationship.
[0119] For candidate decisions with a finite upper bound Only if it satisfies all other candidates simultaneously
[0120]
[0121] At that time, by the candidate The upper bound of the finite candidate, the lower bound of all other candidates, and the comparison relations that hold simultaneously constitute the winning certificate, and this certificate is used to determine the winning candidate. A strict relation is applied to smaller numbers to exclude them from comparison with... A tie, with the larger number treated as a non-strict relation to allow for... Win when the score is tied.
[0122] because When the above relationship holds, for any number less than All candidates have For any number greater than All candidates have Therefore, candidates The results are consistent with the minimum PAPR and minimum numbering priority results under the strict exhaustive definition.
[0123] If no candidate has yet formed the winning certificate, candidates are continued to be scheduled according to the lexicographical key consisting of the candidate's lower bound and candidate number, and a previously unvisited exact replica point of the target mesh is generated for each selected candidate. In the absence of a winning certificate, all candidates selected according to the lexicographical key have not yet been generated. The candidate has a precise replica of the target grid; otherwise, the upper and lower bounds of the candidate are equal to the peak-to-average power ratio of its true target grid, and will automatically satisfy the above unified comparison relationship based on its smallest lexicographical key.
[0124] The total number of all candidates is only Each non-terminating advance adds one generated sample point to the target grid, thus accumulating a maximum of [number] generated samples. The winning candidate is determined by the number of sample points.
[0125] V. On-demand generation of precise replica points and realization of candidate states
[0126] To perform the exact search request for target mesh replica points described in Sections 2 through 4, the natural time index is written as...
[0127]
[0128] Fixed remainder class Corresponding to a length of The inverse discrete Fourier transform, whose input is
[0129]
[0130] For each candidate, a pauseable and resumable inverse transform computation state is established. In the current implementation, this state consists of a fixed-path candidate chain, a candidate evaluation state, and a lazy radix-2 inverse transform graph stored according to remainder classes. The candidate chain stores the current position in the common access order; the candidate evaluation state stores the number of sample points generated, the generated tags, and the current maximum sample point power; the remainder class graph stores the node values and valid tags of the calculated butterfly nodes. These components collectively represent the actual computation progress of the same candidate in the current data block processing flow.
[0131] The state also includes: an identifier indicating that the remainder class has been started; and a natural time index. Remainder Class Remainder class index and terminal node location The mapping relationship; and the target mesh sample count used to determine whether a candidate is complete. The storage format of the above information is not limited, as long as the original calculation progress of the candidate can be restored in the same selection-complete process.
[0132] The first request belongs to the remainder class. When sampling points of the target mesh, the lazy radix-2 inverse transform graph of the remainder class is initiated with the input, and the graph is included in the candidate state by the remainder class identifier; the same graph is restored when sampling points belonging to the same remainder class are requested subsequently. Valid markers in the inverse transform graph are used to distinguish between nodes that have been computed and nodes that have not yet been computed. If the target terminal node and its dependent nodes already have valid cached values, they are read directly; otherwise, only necessary nodes and butterfly patterns that have not yet been valid are computed along the dependencies of the target node, and the node values and valid markers are written back.
[0133] Read the terminal node and press Normalization yields precise complex sampling points for candidate evaluation. Then, the access location, generated tags, generated quantity, and current maximum sample power are updated. Already generated indexes are no longer generated, and each candidate's cache node is stored within its current candidate state. Therefore, the state generated by a candidate advancement can be directly reused by the aforementioned global scheduling or by subsequent winning candidate completion.
[0134] VI. Winning Candidate State Reuse Completion and Natural Time Index Assembly
[0135] Once the winning candidate is determined, the advancement of other candidates is stopped, and the aforementioned candidate chain, candidate evaluation state, and the already initiated remainder-type lazy radix-2 inverse transform graph of the winning candidate are retained from the candidate state set. This process does not regenerate the winning candidate input sequence, does not establish a second set of winning candidate inverse transform states, does not regenerate existing complex points, and does not recalculate butterfly nodes that already have valid labels.
[0136] make The number of exact replicates generated for the winning candidate when selection stops, where .when At that time, starting from the original access location, generate the unvisited locations sequentially along the common access order. Each sample point; the completion process can still calculate necessary nodes that were not previously effective for missing sample points. When The completion loop executes zero times, without generating new target mesh replica points or new radix-2 butterfly nodes. Therefore, the completion phase only processes the target mesh exact replica points that the winning candidate has not yet visited; already generated replica points are not requested again, and other candidates are not advanced.
[0137] All of the winning candidates After the multiple sample points are generated, according to The natural chronological order, with , and Locate the cached terminal node, read the node, and press... Normalized write position This forms a complete inner-layer complex waveform that serves as the input to the AFDM emitter chain. The completion is skipped, but the natural time index assembly is still performed. The assembly only includes index positioning, cache reading, normalization, and array writing, and does not trigger new radix-2 butterfly calculations, butterfly multiplications, or butterfly additions.
[0138] VII. Fixed Experimental Examples
[0139] This embodiment uses the following configuration: , , , QPSK modulation, using a total of 256 fixed data blocks; each candidate contains The inner layer target mesh is repeated at a certain number of points. This method uses the same input, codebook, energy normalization, target mesh, and flat rules as the strict exhaustive method.
[0140] For each data block, a complete exhaustive search is performed to generate all possible results. The method generates precise sample points on demand and determines the optimality of the same candidate set based on the upper and lower bounds of the candidates. After determining the winning candidate, the inverse transformation of the same candidate is reused to fill in the missing sample points, forming a complete inner complex waveform containing all 2048 target grid sample points.
[0141] The block-by-block comparison results show that the winning candidate numbers obtained by this method on 256 data blocks are consistent with the strict exhaustive search. The complete inner complex waveform and its PAPR after state multiplexing and completion are also consistent with the results of the strict exhaustive search within the numerical accuracy range. The winning certificates formed all satisfy the candidate upper and lower bound comparison relationship.
[0142] Of the 256 data blocks, 22 blocks generated winning certificates before the winning candidate had generated all 2048 sample points, and continued to complete the missing sample points from the same retained candidate state; the remaining 234 data blocks had a complete winning candidate when the candidate selection stopped. On average, the winning candidate had generated 1972.637 sample points when the selection stopped, and an average of 75.363 sample points were generated during the completion phase, totaling 2048. This indicates that the completion process only generated target mesh repeat sample points that the winning candidate had not yet visited.
[0143] This embodiment separately counts the number of multiple multiplications, multiple additions, and the number of target grid sample points generated. The statistical results of this method include two consecutive stages: candidate selection and successful candidate state reuse and completion. The average counts for the above three items using strict exhaustive enumeration are 147456, 229376, and 32768, respectively. The corresponding average counts for this method are 17116.559, 25962.301, and 3571.316, respectively, representing reductions of 88.39%, 88.68%, and 89.10%.
[0144] Figure 2 Normalizing the average counts of the strict exhaustive search on the above three indicators to 100%, the corresponding relative counts of this method are 11.61%, 11.32%, and 10.90%, respectively. The experimental results show that this method can reduce the number of complex multiplications, complex additions, and target grid duplicate points generated while maintaining consistency with the strict exhaustive search for winning candidates and complete winning waveforms on a fixed target grid. These experimental results correspond to the fixed parameter configuration of this embodiment and do not constitute a limitation on the computational cost reduction under other parameter configurations.
[0145] The method and system of this invention can be deployed in the transmitter baseband processor of an AFDM communication system, implemented in software, hardware, or a combination of both. For software implementation, data such as candidate states, inverse transform graph buffers, and upper and lower bounds can be stored in the transmitter's random access memory, with corresponding calculation steps executed by a digital signal processor or application-specific integrated circuit (ASIC). For hardware implementation, each module can be implemented by corresponding logic circuits in a field-programmable gate array (FPGA) or ASIC to meet the processing latency and throughput requirements of real-time communication systems. All of the above implementation methods are within the protection scope of this invention.
[0146] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A method for generating low peak-to-average power ratio transmit waveforms in simulated radio frequency demultiplexing (AFDM) based on precise candidate upper and lower bound search, characterized in that, Includes the following steps: (a) Obtain the complex data symbol block to be sent, construct a finite number of AFDM chirp selection mapping candidates for the data symbol block, define the inner layer complex sampling point and the peak-to-average power ratio of the target grid for each candidate on a fixed target grid, and stipulate that when the peak-to-average power ratio of the target grid is equal, the candidate with the smaller number is selected. (b) For each candidate, maintain a candidate lower bound that is not higher than the peak-to-average power ratio of the candidate's true target grid and a candidate upper bound that is not lower than the peak-to-average power ratio of the true target grid, set the initial mathematical values of the candidate lower bound and the candidate upper bound, and specify the target grid sample access order for all candidates to use. (c) Establish candidate inverse transformation calculation states that can be paused and resumed for each candidate, generate accurate composite sample points of the target grid as needed according to the target grid sample point access order, and reuse the completed inverse transformation calculations; (d) Schedule the candidates according to the dictionary key consisting of the candidate lower bound and the candidate number, generate a target grid exact sample point that has not been visited before, and update the candidate lower bound and candidate upper bound of the selected candidate according to the generated sample points; (e) Determine whether the winning candidate can be determined based on the comparison between the upper bound of each candidate and the lower bound of all other candidates; If it cannot be determined, return to step (d) to continue generating sample points and updating the upper and lower bounds until a winning candidate is determined; (f) After determining the winning candidate, stop advancing other candidates, retain the candidate inverse transformation calculation state of the winning candidate, and only continue to generate target mesh exact replica points that have not yet been visited from the retained state, without reconstructing the winning candidate, without regenerating existing sample points, and without recalculating cached nodes; (g) After the winning candidate generates all the target mesh exact replica points, they are arranged in natural time index order to form a complete inner layer complex waveform.
2. The method according to claim 1, characterized in that, In step (a), the length of the data symbol block is determined. and oversampling factor The determined length is On a fixed target grid, the peak-to-average power ratio of each candidate inner-layer complex sample point and the target grid is defined, and the complex data symbol block is represented as... ,in, , Set the oversampling factor , and The number of candidates is a positive integer. Not less than An integer, given a finite real base second chirp parameter And for candidate numbers ,according to structure AFDM chirp selection mapping candidates; let in, and the first A candidate inner layer replica point on the fixed target mesh is defined as... by As the first Peak-to-average power ratio of the target grid for each candidate.
3. The method according to claim 2, characterized in that, In step (b), the lower bound of each candidate... and candidate upper bound satisfy The initial mathematical value of the candidate lower bound is The initial mathematical value of the candidate upper bound is positive infinity; the target grid sample point access order is as follows: in, Indicates to The bitwise binary representation performs bit reversal, and the access sequence forms a complete uniform subgrid at a preset milestone. ,satisfy So that the access order is before Each position constitutes a complete Double uniform subgrid.
4. The method according to claim 3, characterized in that, In step (c), the natural time index of the target grid is decomposed into ,in, , Make each remainder class Corresponding to a length of The inverse discrete Fourier transform; only when the first request belongs to the remainder class. When the target mesh is accurately replicated, the inert radix-2 inverse transform graph of the remainder class is initiated, and a sequence is generated. The length of this remainder class is The inverse discrete Fourier transform input is used; when subsequent requests are made for exact sample points of the target mesh belonging to the same remainder class, the same inert radix-2 inverse transform graph is restored and the butterfly nodes already calculated therein are reused; the candidate intra-inverse transform calculation state includes the current position in the target mesh sample point access order, the identifier of the activated remainder class, the calculated butterfly nodes of each activated remainder class, and the natural time index. With the corresponding remainder class Remainder class index The mapping relationship between the terminal nodes of the inert radix-2 inverse transform graph, the generated labels of the target grid's accurate complex points, and the currently observed maximum sample power.
5. The method according to claim 4, characterized in that, In step (d), after the selected candidates have been generated After accurately replicating each target grid point, according to Update the lower bound of the selected candidates.
6. The method according to claim 5, characterized in that, In step (d), when the number of generated sample points reaches a non-final preset milestone, the candidate upper bound of the candidate is updated using the precise complex sample points on the corresponding complete uniform subgrid; when the number of generated sample points reaches the final preset milestone... When the candidate's lower bound and upper bound are updated to their true target grid peak-to-average power ratio; Wherein, when the number of samples generated by the selected candidate according to the target grid sample point access order reaches the complete Preset milestones corresponding to double uniform subgrids season And in Time calculation exist season The candidate upper bound of the selected candidate is updated with the smaller of the currently saved candidate upper bound and the candidate upper bound calculated in this operation.
7. The method according to claim 6, characterized in that, In step (e), any candidate with a finite upper bound is taken as a candidate to be decided. When the upper bound of the proposed candidate is strictly less than the lower bound of all other candidates whose numbers are less than the proposed candidate, and is less than or equal to the lower bound of all other candidates whose numbers are greater than the proposed candidate, the proposed candidate is determined as the winning candidate. Among them, for candidates to be determined with a finite upper bound When both conditions are met At that time, candidates The candidate is determined as the winning candidate; the finite number of candidates is denoted as... Each time a candidate is advanced but not yet completed, only one exact replica point of the target mesh that has not been visited before is generated. Each candidate generates at most one exact replica point of the target mesh. The aforementioned sample points, and at the latest generated by accumulating all candidates. The winning candidate is determined when the number of the aforementioned sample points is reached.
8. The method according to claim 7, characterized in that, In steps (f) and (g), let This represents the number of exact replica points of the target mesh that the winning candidate has generated when candidate selection stops, where, ;when At that time, the inverse transformation within the candidate domains of the retained winning candidates is used to calculate the state, generating only those that have not yet been visited. Each target grid is accurately replicated at multiple points; when At that time, no new target mesh exact replica points or new radix-2 butterfly nodes will be generated; in all of the winning candidates After generating accurate replica points for each target grid, according to The order is determined by and Determine the cached terminal nodes in the lazy radix-2 inverse transform graph of the corresponding remainder class, read the terminal nodes and follow the... Normalization is performed to assemble the complete inner complex waveform; the reading, normalization, and assembly do not trigger new radix-2 butterfly node calculations, butterfly complex multiplications, or butterfly complex additions.
9. A low peak-to-average power ratio (PAPR) transmit waveform generation system for AFDM based on precise candidate upper and lower bound search, characterized in that, include: The candidate construction module is used to construct a finite number of AFDM chirp selection mapping candidates. It defines the inner layer complex points of each candidate and the peak-to-average power ratio of the target grid on a fixed target grid, and stipulates that when the peak-to-average power ratios of the target grids are equal, the candidate with the smaller number is selected. The upper and lower bound maintenance module is used to maintain the candidate lower bound and candidate upper bound for each candidate, set the initial mathematical values of the candidate lower bound and candidate upper bound, and specify the target grid sample point access order for all candidates to use. The inertial inverse transformation calculation module is used to establish candidate intra-candidate inverse transformation calculation states that can be paused and resumed for each candidate, generate accurate composite sample points of the target grid as needed according to the target grid sample point access order, and reuse the completed inverse transformation calculations. The scheduling and lower bound update module is used to schedule candidates according to the dictionary key consisting of the candidate lower bound and the candidate number, generate a target grid accurate resample point that has not been visited before, and update the candidate lower bound of the selected candidate based on the generated sample points. The upper bound update module is used to update the upper bound of the selected candidates based on the generated sample points. The optimality determination module is used to determine the winning candidate based on the comparison relationship between the upper bound of each candidate and the lower bound of all other candidates. If it cannot be determined, it triggers the generation of more samples and updates the upper and lower bounds until the winning candidate is determined. The state reuse and completion module is used to retain the inverse transform calculation state within the candidate after the winning candidate is determined, and to continue generating the target mesh exact replica points that have not yet been visited only from the retained state; The waveform assembly output module is used to arrange all target grid accurate replica points in natural time index order after the winning candidate generates all the target grid accurate replica points to form a complete inner layer complex waveform.
