A Nesterov gradient accelerated Markov chain Monte Carlo detection method
By introducing Nesterov gradient acceleration technology in Markov chain Monte Carlo detection, the problem of inefficient search space in ultra-large-scale MIMO systems is solved, and efficient detection methods are realized, approximate the optimal maximum likelihood detection is achieved. performance.
Patent Information
- Application Number
- CN202211613554.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-12-15
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2042-12-15
AI Technical Summary
The existing Markov chain Monte Carlo detection technology in ultra-large-scale MIMO systems has greatly increased the number of random walking steps due to inefficient search space, which hinders the application of this technology.
Using Nesterov gradient acceleration technology, Nesterov gradient descent method is introduced during the random walk of Markov chain Monte Carlo, to improve search efficiency and reduce invalid search attempts.
It significantly improves the efficiency of Markov chain Monte Carlo detection, can approximate the performance of optimal maximum likelihood detection, and reduces computing overhead, and is suitable for ultra-large-scale MIMO systems.
Smart Images

Figure CN115987746B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a Nesterov gradient accelerated Markov chain Monte Carlo detection method, belonging to the technical field of wireless communication. Background Art
[0002] In recent years, Multiple-Input Multiple-Output (MIMO) technology has gradually evolved towards ultra-large-scale MIMO, which supports the explosive growth of data traffic demand in the sixth-generation mobile communication system by deploying more antennas at base stations and wireless access points to expand the capacity of wireless networks exponentially. However, the improvement of system spectrum and power efficiency depends on efficient detection schemes at the receiving end. In traditional detection schemes, the complexity of maximum likelihood detection with the best performance increases exponentially with the number of decision variables, making it impractical; low-complexity linear detection schemes are criticized for their poor performance. Achieving an ideal compromise between complexity and receiving performance has become the core of attention in the design of detection schemes.
[0003] Markov Chain Monte Carlo is a statistical machine learning technique based on random approximation. It can accurately infer the target probability distribution with moderate complexity and has received extensive attention in the design of MIMO detection schemes. This technique extracts samples from a simple proposed distribution to form a Markov chain that converges asymptotically to a unique stationary distribution. In the MIMO detection problem, the stationary distribution is set as the posterior probability distribution of the transmitted symbols, and the posterior mean estimate can be approximated by sampling and statistical averaging.
[0004] However, most of the existing Markov Chain Monte Carlo detection technologies are based on random walks to find the optimal solution, which is very inefficient in exploring the search space. As the search space increases, the number of random walks required will increase significantly, hindering the application of this technology in ultra-large-scale MIMO systems. Therefore, improving search efficiency is an urgent problem to be solved in the practical application of Markov Chain Monte Carlo detection technology. Summary of the invention
[0005] Technical problem: In view of the above-mentioned deficiencies in the prior art, the present invention provides a Markov chain Monte Carlo detection method with Nesterov gradient acceleration, which uses Nesterov accelerated gradient descent technology to improve the search efficiency of Markov chain Monte Carlo random walk, reduce a large number of invalid search attempts, and achieve performance close to the optimal maximum likelihood detection with a lower computational overhead.
[0006] Technical solution: A Nesterov gradient accelerated Markov chain Monte Carlo detection method of the present invention comprises the following steps:
[0007] S1, initialize P parallel samplers and set the sampling iteration variable i=1;
[0008] S2, calculating the proposal vector of each of the P parallel samplers, including the Nesterov gradient proposal and the random walk noise; performing the Metropolis-Hastings acceptance step, obtaining the sample of each sampler, and setting the sampling iteration variable i=i+1;
[0009] S3, judging whether the sampling iteration variable i is greater than the total sampling iteration number N, if not, executing step S4; if so, giving a judgment detection result, and the method ends;
[0010] S4. Determine whether the early stopping condition is met. If not, jump to step S2; if so, give a judgment detection result and the method ends.
[0011] in,
[0012] The step S1 specifically includes:
[0013] S101, calculate the Hessian matrix G = H H H, matched filter vector and the Lipschitz constant for Nesterov gradient descent Where H is the channel matrix, y is the received signal vector, (·) H represents the conjugate transpose of a matrix, ||·|| F represents the Frobenius norm of the matrix;
[0014] S102, calculate the inverse of the Hessian matrix (H H H) -1 , as the covariance of the proposal vector, where (·)- 1 Represents the inverse of the matrix, the inverse of the Hessian matrix (H H H) -1 Perform Cholesky decomposition and normalize each row to obtain the covariance decomposition lower triangular matrix L;
[0015] S103, randomly initialize the initial estimation vector z0 of each of the P parallel samplers; quantize the elements in each dimension of z0 to orthogonal amplitude modulation constellation points to obtain the initial sample x0=Q(z0) of each sampler, where Q(·) represents the quantization operation of mapping a complex value to the orthogonal amplitude modulation constellation point with the closest Euclidean distance to it; calculate the residual r0=y-Hx0 of the initial sample;
[0016] S104 , setting the sampling iteration variable i=1.
[0017] The step S2 specifically includes, for each of the P parallel samplers:
[0018] S201. Use the Nesterov gradient descent method to obtain the gradient proposal z i , as the proposed mean;
[0019] S202, generate a Gaussian random vector v with zero mean and unit variance, and calculate sLv to obtain random walk noise, where s is the step size of the random walk noise, and L is the lower triangular matrix of the covariance decomposition, so that the proposed covariance is the inverse of the Hessian matrix;
[0020] S203, calculate the proposal vector z p , the expression is:
[0021] z p =z i +sLv;
[0022] S204, propose vector z p Quantize to get candidate sample x p =Q(z p ), where Q(·) represents the quantization operation of mapping a complex value to the orthogonal amplitude modulation constellation point with the closest Euclidean distance to it; the residual r of the candidate sample is calculated p =y-Hx p , where y is the received signal vector and H is the channel matrix;
[0023] S205. Calculate the acceptance probability α of the candidate sample according to the Metropolis-Hastings criterion, expressed as:
[0024]
[0025] Among them, min means taking the minimum value, ||·|| means the Euclidean norm of the vector, and r i-1 is the sample residual of the previous sampling iteration;
[0026] S206, take a random number τ uniformly distributed between (0, 1), determine whether α is less than τ, if not, let the current sample x i =x p ; If so, discard the candidate sample x p , let the current sample x i =x i-1 , where x i-1 is the sample of the previous sampling iteration;
[0027] S207, calculate the residual r of the current sample i =y-Hx i , update the random walk noise step size s, the expression is:
[0028] s=max(d,k||r i ||)
[0029] Wherein, d is half of the minimum Euclidean distance between transmitted orthogonal amplitude modulation constellation symbols, k is a preset positive number, and max indicates the maximum value;
[0030] S208. Set the sampling iteration variable i=i+1.
[0031] The Nesterov gradient descent method described in step S201 specifically includes:
[0032] S211, set the gradient descent starting point z g,0 =x i-1 , where x i-1 is the sample of the previous sampling iteration; set the gradient descent count variable t = 1;
[0033] S212, superposition momentum term, the expression is:
[0034]
[0035] Among them, z g,t-1 is the gradient update vector in the previous round of gradient descent, β is the momentum factor, Δz is the momentum term, is the gradient update vector with the momentum term superimposed;
[0036] S213, Calculation The gradient direction at The expression is:
[0037]
[0038] Among them, z g,t is the gradient update vector in the current gradient descent, is the Lipschitz constant of Nesterov gradient descent, used for the gradient descent step size, H is the channel matrix, (·) H represents the conjugate transpose of the matrix, y is the received signal vector, is the matched filter vector, G is the Hessian matrix;
[0039] S214. Update the momentum term, the expression is:
[0040] Δz=z g,t -z g,t-1 ;
[0041] S215, let t = t + 1, determine whether t is greater than the total number of gradient descents N g If not, jump to step S212; if yes, let the gradient propose in, is the counting variable t equal to N g The gradient update vector at .
[0042] The method for determining whether the early stopping condition is met in step S4 specifically includes:
[0043] S401, determining the sample with the smallest residual norm of each sampler of P parallel samplers up to the current sampling iteration, to obtain P samples;
[0044] S402: Determine the sample x with the smallest residual norm among the P samples obtained in step S401. b ;
[0045] S403, statistical sample x b The number of occurrences N in the P samples obtained in step S401 b ; Calculate sample x b The residual norm I b =||y-Hx b ||, where y is the received signal vector, H is the channel matrix, and ||·|| represents the Euclidean norm of the vector;
[0046] S404, judgment condition N b >P / 2 and Whether it satisfies, where η is a preset coefficient, represents the expectation operation, and n is the additive white Gaussian noise vector.
[0047] The samples obtained in step S2 also include: quantizing z g,t The obtained samples, where z g,t is the gradient update vector in the current gradient descent in step S213.
[0048] The samples obtained in step S2 constitute a sample list, and the decision detection result described in step S3 or S4 is a hard decision detection result of selecting a sample with the smallest residual norm from the sample list, or a soft decision detection result of calculating a log-likelihood ratio from the sample list.
[0049] Beneficial effects: Compared with the prior art, the technical solution provided by the present invention has the following beneficial effects:
[0050] The present invention closely integrates the accelerated Nesterov gradient descent scheme with the Markov chain Monte Carlo random walk process. On the one hand, gradient descent is used to provide efficient guidance for random walks, which significantly improves the efficiency of the sampling process and the detection method, and can better cope with the high-dimensional space search problem presented in ultra-large-scale MIMO systems; on the other hand, random walks can effectively correct the problem that gradient descent in continuous space cannot converge to the optimal solution in the discrete search space of MIMO detection. Compared with existing linear detection and nonlinear detection schemes based on message passing, significant performance gains can be achieved. The present invention adopts a first-order gradient descent scheme that is friendly to hardware implementation. Compared with the closest existing technology, it fully reduces the number of complex operations such as matrix inversion, maintains a low computational complexity, and can be well extended to the detection problem of ultra-large-scale MIMO systems. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] Figure 1 It is a flow chart of the Nesterov gradient accelerated Markov chain Monte Carlo detection method according to an embodiment of the present invention;
[0052] Figure 2 A system block diagram of a Nesterov gradient accelerated Markov chain Monte Carlo detection method according to an embodiment of the present invention;
[0053] Figure 3 The figure is a comparison chart of the bit error rate (BER) performance between the method of the present invention and the comparative method. DETAILED DESCRIPTION
[0054] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.
[0055] 1. System Model Used in This Example
[0056] Consider a system with N t Transmitting antennas and N r In this embodiment, the transmitter uses quadrature amplitude modulation (QAM), and the constellation symbol set is recorded as The average energy is normalized to 1. The transmission bit stream is modulated by QAM to obtain the transmission symbol vector After transmission through the wireless channel, the received signal y can be expressed as:
[0057] y=Hx+n
[0058] in, is the channel gain matrix, n is zero mean and variance is σ 2The receiver signal-to-noise ratio (SNR) is defined as in, represents the expected operation, and ||·|| represents the Euclidean norm of the vector. In this embodiment, it is assumed that the receiving end perfectly knows the channel state information H and the noise variance σ 2 .
[0059] 2. Specific steps of this embodiment
[0060] like Figure 2 As shown, an embodiment of the present invention provides a system block diagram of a Nesterov gradient accelerated Markov chain Monte Carlo detection method, which includes P parallel Markov chain Monte Carlo samplers. First, each sampler is initialized to obtain an initial sample x0. Then, each sampler completes N sampling times to obtain samples x1, x2, ..., x N , N is the total number of sampling iterations. Each sampling includes Nesterov gradient descent, adding random walk noise, quantization, and Metropolis-Hastings (MH) acceptance steps. Finally, the sampled samples are merged to obtain a sample list, and the decision is performed based on the sample list. Figure 1 As shown, an embodiment of the present invention provides a flow chart of a Markov chain Monte Carlo detection method based on Nesterov gradient descent,
[0061] S1, initialize P parallel samplers, set sampling iteration variable i = 1;
[0062] S2, calculating the proposal vector of each of the P parallel samplers, including the Nesterov gradient proposal and the random walk noise; performing the Metropolis-Hastings acceptance step, obtaining the sample of each sampler, and setting the sampling iteration variable i=i+1;
[0063] S3, judging whether the sampling iteration variable i is greater than the total sampling iteration number N, if not, executing step S4; if so, giving a judgment detection result, and the method ends;
[0064] S4. Determine whether the early stopping condition is met. If not, jump to step S2; if so, give a judgment detection result and the method ends.
[0065] Specifically:
[0066] S1. Initialize P parallel Markov chain Monte Carlo samplers, specifically including the following steps:
[0067] S101, calculate the Hessian matrix G = H HH, matched filter vector and the Lipschitz constant for Nesterov gradient descent For the Nesterov gradient descent process, where (·) H represents the conjugate transpose of a matrix, ||·|| F represents the Frobenius norm of the matrix;
[0068] S102, calculate the inverse of the Hessian matrix (H H H) -1 , then (H H H) -1 The lower triangular matrix L is obtained by performing Kholesky decomposition and normalizing the decomposition matrix row by row, which is used to construct the covariance of the random walk noise. This makes the correlation between the random walk noises consistent with the correlation between the noise streams observed from the perspective of the transmitted symbol vector, thereby improving the search effect of the random walk. At the same time, normalization ensures the independence between the random walk noise and the channel gain.
[0069] S103, initialize the initial estimation vector z0 of each of the P parallel samplers. In this embodiment, a random initialization method is adopted; quantize the elements on each dimension of z0 to QAM constellation points to obtain the initial sample x0=Q(z0) of each sampler, where Q(·) represents the quantization operation of mapping a complex value to the QAM constellation point with the closest Euclidean distance to it; calculate the residual r0=y-Hx0 of the initial sample;
[0070] S104 , setting the sampling iteration variable i=1.
[0071] S2. Calculate the proposal vector of each of the P parallel samplers, including the Nesterov gradient proposal and the random walk noise, specifically including the following steps:
[0072] S20l, use the Nesterov gradient descent method to obtain the gradient proposal z i , as the proposed mean;
[0073] S202, generate a Gaussian random vector v with zero mean and unit variance, calculate sLy to obtain random walk noise, where s is the step size of the random walk noise, and L is the covariance decomposition lower triangular matrix, so that the proposed covariance is the inverse of the Hessian matrix, satisfying the correlation described in step S102;
[0074] S203, calculate the proposal vector z p , the expression is:
[0075] z p =z i +sLv
[0076] Furthermore, the Nesterov gradient descent method includes a continuous N g times of accelerated gradient descent, where N g is the total number of gradient descents, so that the estimated residual can be fully reduced and a good balance is achieved with the effect of random walk, including:
[0077] S211, set the gradient descent starting point z g,0 =x i-1 , where x i-1 is the sample of the previous sampling iteration; set the gradient descent count variable t = 1;
[0078] S212, superposition momentum term, the expression is:
[0079]
[0080] Among them, z g,t-1 is the gradient update vector in the previous round of gradient descent, β is the momentum factor, which is taken as 0.9 in this embodiment, and Δz is the momentum term. is the gradient update vector with the momentum term superimposed. By superimposing the momentum term, the average gradient update of multiple gradient descents is used to replace the current gradient update, thereby accelerating the initial stage of multiple gradient descents and slowing down the oscillation and enhancing stability in the later stage.
[0081] S213, Calculation The gradient direction at Superimposed on The expression is:
[0082]
[0083] Among them, z g,t is the gradient update vector in the current gradient descent, is the Lipschitz constant of Nesterov gradient descent, used for the gradient descent step size, is the matched filter vector, G is the Hessian matrix;
[0084] S214. Update the momentum term, the expression is:
[0085] Δz=z g,t -z g,t-1 ;
[0086] S215, let t = t + 1, determine whether t is greater than the total number of gradient descents N g If not, jump to step S212; if yes, let the gradient propose in, is the counting variable t equal to N g The gradient update vector at time ;
[0087] Different from the high-order gradient descent schemes including Newton's method used in the prior art, the Nesterov gradient descent method described in the present invention belongs to a first-order gradient descent scheme, which avoids the complex matrix inversion operation in the high-order gradient descent scheme and has lower computational complexity. Therefore, the present invention can better adapt to ultra-large-scale MIMO systems.
[0088] Further, the MH acceptance step is performed to obtain candidate samples based on the proposal vector, and the candidate samples are accepted or rejected according to appropriate criteria, thereby obtaining samples of each sampler, which specifically includes the following steps:
[0089] S204: the proposed vector z described in step S203 p Quantize to get candidate sample x p =Q(z p ), where Q(·) represents the quantization operation of mapping a complex value to the QAM constellation point with the closest Euclidean distance; the residual r of the candidate sample is calculated p =y-Hx p ;
[0090] S205. Calculate the acceptance probability α of the candidate sample according to the MH criterion. In this embodiment, assuming that the transmitted symbols are equally likely a priori, the acceptance probability α is the ratio of the likelihood probability of the candidate sample to the likelihood probability of the sample in the previous sampling iteration, expressed as:
[0091]
[0092] Among them, min means taking the minimum value, r i-1 is the sample residual of the previous sampling iteration;
[0093] S206, take a random number τ uniformly distributed between (0, 1), determine whether α is less than τ, if not, let the current sample x i =x p ; If so, discard the candidate sample x p , let the current sample x i =x i-1 , where x i-1 is the sample of the previous sampling iteration;
[0094] S207, calculate the residual r of the current sample i =y-Hx i , update the random walk noise step size s, the expression is:
[0095] s=max(d,k||r i ||)
[0096] Where d is half of the minimum Euclidean distance between transmitted QAM constellation symbols, and k is a preset positive number, which is taken as 1 / N in this embodiment. t , max means taking the maximum value; when the Euclidean distance between the sample and the maximum likelihood solution is far, k||r i || takes a larger value, the corresponding random walk step length is larger, and the search space can be explored more comprehensively; when the Euclidean distance between the sample and the maximum likelihood solution is close, k||r i || takes a smaller value, and the corresponding random walk step size is smaller, so a fine search can be performed near the maximum likelihood solution. Limiting the minimum step size to d can avoid the flameout problem existing in the existing Markov chain Monte Carlo technology, so that the sampler can jump out of the local optimal solution;
[0097] S208. Set the sampling iteration variable i=i+1.
[0098] S3, judging whether the sampling iteration variable i is greater than the total sampling iteration number N, if not, executing step S4; if yes, making a judgment;
[0099] S4: Determine whether the early stopping condition is met. If not, jump to step S2; if so, make a decision.
[0100] Furthermore, an early stopping condition is set to dynamically adjust the total number of iterations to avoid invalid searches and reduce computational complexity, specifically including the following steps:
[0101] S401, determining the sample with the smallest residual norm of each sampler of P parallel samplers up to the current sampling iteration, to obtain P samples;
[0102] S402: Determine the sample x with the smallest residual norm among the P samples obtained in step S401. b ;
[0103] S403, statistical sample x b The number of occurrences N in the P samples obtained in step S401 b ; Calculate sample x b The residual norm l b =||y-Hx b ||;
[0104] S404, judgment condition N b >P / 2 is satisfied, that is, sample x b Whether the frequency of occurrence in the P parallel samplers is high enough; Whether it satisfies, where η is a preset coefficient, which is taken as 1 in this embodiment, that is, sample x b Is the residual norm of small enough? If the condition N b >P / 2 and If it satisfies, then the sample x b The probability of taking the maximum likelihood solution is high enough to allow early stopping.
[0105] Furthermore, the samples obtained in step S2 also include: quantizing z g,t The obtained samples, where z g,t It is the gradient update vector in the current gradient descent in step S213, and additional samples are obtained in the Nesterov gradient descent method to improve the decision performance.
[0106] Furthermore, the samples obtained in step S2 constitute a sample list The decision in step S3 or S4 is a hard decision of selecting a sample with the smallest residual norm from the sample list, or a soft decision of calculating the log-likelihood ratio from the sample list. In this embodiment, the hard decision is executed to obtain the detection result The expression is:
[0107]
[0108] Among them, arg min represents the minimum point of the function.
[0109] 3. Implementation Effect
[0110] In order to enable those skilled in the art to better understand the solution of the present invention, the following provides a comparison of the results of the Nesterov gradient accelerated Markov chain Monte Carlo detection method and the comparison method in this embodiment.
[0111] The simulation scenario considered is described as follows:
[0112] The uplink MIMO transmission link in the urban macrocell non-line-of-sight scenario uses 16-QAM constellation modulation. The base station is equipped with 32 dual-polarization antennas, corresponding to the number of receiving antennas N. r =64, the antenna spacing is half a wavelength, the base station height is 25 meters, the cellular sector covers 60°, the radius is 500 meters, and 8 single-antenna users are randomly placed in the sector, corresponding to the number of transmitting antennas N t =8, the protection distance between the user and the base station is 10 meters, and each user moves along a straight line at a speed of 1 meter per second. The carrier frequency of the uplink MIMO transmission link is 2.53 GHz, the channel bandwidth is 20 MHz, and the number of effective subcarriers is 256.
[0113] Figure 3A comparison chart of the BER performance of the present invention and the comparison method in the simulation scenario is shown, and the comparison method includes the expected propagation detection method in the document “Expectation propagation detection for high-order high-dimensional MIMOsystems, IEEE Trans. Commun., vol. 62, no. 8, pp. 2840–2849, Aug. 2014.” (hereinafter referred to as detection scheme 1), the detection method combining Newton gradient descent method and Markov chain Monte Carlo sampling in the document “Metropolis-Hastings random walk along the gradient descent direction for MIMO detection, in Proc. IEEE Int. Conf. Commun. (ICC), Montreal, QC, Canada, Jun. 2021, pp. 1–7.” (hereinafter referred to as detection scheme 2), and the theoretically optimal maximum likelihood detection BER lower bound. The number of iterations of detection scheme 1 is T=10, the number of parallel samplers used in detection scheme 2 is P=16, and the two BER-SNR curves correspond to the effects of total iteration numbers N=8 and N=16 respectively. The number of parallel samplers of the present invention is also taken as P=16. Under the configuration of an average number of iterations of N=5.1, a performance gain is achieved relative to detection scheme 2 with a larger number of iterations. At the same time, the performance of the present invention is significantly better than that of detection scheme 1, and is close to the maximum likelihood BER lower bound.
[0114] What is described above in conjunction with the accompanying drawings is only a preferred embodiment of the present invention and cannot be used to limit the scope of rights included in the present invention. It should be understood that any equivalent changes made without departing from the spirit of the present invention are within the scope of protection covered by the claims of the present invention.
Claims
1. A Nesterov gradient accelerated Markov chain Monte Carlo detection method, characterized in that: The method comprises the following steps: S1, initialize P parallel samplers and set the sampling iteration variable i=1; S2, calculating the proposal vector of each of the P parallel samplers, including the Nesterov gradient proposal and the random walk noise; performing the Metropolis-Hastings acceptance step, obtaining the sample of each sampler, and setting the sampling iteration variable i=i+1; S3, judging whether the sampling iteration variable i is greater than the total sampling iteration number N, if not, executing step S4; if so, giving a judgment detection result, and the method ends; S4, judging whether the early stopping condition is met, if not, jumping to step S2; if so, giving the judgment detection result, and the method ends; in, The step S1 specifically includes: S101, calculate the Hessian matrix G = H H H, matched filter vector and the Lipschitz constant for Nesterov gradient descent Where H is the channel matrix, y is the received signal vector, (·) H represents the conjugate transpose of a matrix, ||·|| F represents the Frobenius norm of the matrix; S102, calculate the inverse of the Hessian matrix (H H H) -1 , as the covariance of the proposal vector, where (·) -1 Represents the inverse of the matrix, the inverse of the Hessian matrix (H H H) -1 Perform Cholesky decomposition and normalize each row to obtain the covariance decomposition lower triangular matrix L; S103, randomly initialize the initial estimation vector z0 of each of the P parallel samplers; quantize the elements in each dimension of z0 to orthogonal amplitude modulation constellation points to obtain the initial sample x0=Q(z0) of each sampler, where Q(·) represents the quantization operation of mapping a complex value to the orthogonal amplitude modulation constellation point with the closest Euclidean distance to it; calculate the residual r0=y-Hx0 of the initial sample; S104, set sampling iteration variable i=1; The step S2 specifically includes, for each of the P parallel samplers: S201. Use the Nesterov gradient descent method to obtain the gradient proposal z i , as the proposed mean; S202, generate a Gaussian random vector v with zero mean and unit variance, and calculate sLv to obtain random walk noise, where s is the step size of the random walk noise, and L is the lower triangular matrix of the covariance decomposition, so that the proposed covariance is the inverse of the Hessian matrix; S203, calculate the proposal vector z p , the expression is: z p =z i +sLv; S204, propose vector z p Quantize to get candidate sample x p =Q(z p ), where Q(·) represents the quantization operation of mapping a complex value to the orthogonal amplitude modulation constellation point with the closest Euclidean distance to it; the residual r of the candidate sample is calculated p =y-Hx p , where y is the received signal vector and H is the channel matrix; S205. Calculate the acceptance probability α of the candidate sample according to the Metropolis-Hastings criterion, expressed as: Among them, min means taking the minimum value, ||·|| means the Euclidean norm of the vector, and r i-1 is the sample residual of the previous sampling iteration; S206, take a random number τ uniformly distributed between (0,1), determine whether α is less than τ, if not, let the current sample x i =x p ; If so, discard the candidate sample x p , let the current sample x i =x i-1 , where x i-1 is the sample of the previous sampling iteration; S207, calculate the residual r of the current sample i =y-Hx i , update the random walk noise step size s, the expression is: s=max(d,k||r i ||) Wherein, d is half of the minimum Euclidean distance between transmitted orthogonal amplitude modulation constellation symbols, k is a preset positive number, and max indicates the maximum value; S208, set sampling iteration variable i=i+1; The Nesterov gradient descent method described in step S201 specifically includes: S211, set the gradient descent starting point z g,0 =x i-1 , where x i-1 is the sample of the previous sampling iteration; set the gradient descent count variable t = 1; S212, superposition momentum term, the expression is: Among them, z g,t-1 is the gradient update vector in the previous round of gradient descent, β is the momentum factor, Δz is the momentum term, is the gradient update vector with the momentum term superimposed; S213, Calculation The gradient direction at The expression is: Among them, z g,t is the gradient update vector in the current gradient descent, is the Lipschitz constant of Nesterov gradient descent, used for the gradient descent step size, H is the channel matrix, (·) H represents the conjugate transpose of the matrix, y is the received signal vector, is the matched filter vector, G is the Hessian matrix; S214. Update the momentum term, the expression is: Δz=z g,t -with g,t-1 ; S215, let t = t + 1, determine whether t is greater than the total number of gradient descents N g If not, jump to step S212; if yes, let the gradient propose in, is the counting variable t equal to N g The gradient update vector at time ; The method for determining whether the early stopping condition is met in step S4 specifically includes: S401, determining the sample with the smallest residual norm of each sampler of P parallel samplers up to the current sampling iteration, to obtain P samples; S402: Determine the sample x with the smallest residual norm among the P samples obtained in step S401. b ; S403, statistical sample x b The number of occurrences N in the P samples obtained in step S401 b ; Calculate sample x b The residual norm l b =||y-Hx b ||, where y is the received signal vector, H is the channel matrix, and ||·|| represents the Euclidean norm of the vector; S404, judgment condition N b >P / 2 and Whether it satisfies, where η is a preset coefficient, represents the expectation operation, and n is the additive white Gaussian noise vector.
2. The Nesterov gradient accelerated Markov chain Monte Carlo detection method according to claim 1, characterized in that: The samples obtained in step S2 also include: quantizing z g,t The obtained samples, where z g,t is the gradient update vector in the current gradient descent in step S213.
3. The Nesterov gradient accelerated Markov chain Monte Carlo detection method according to claim 1, characterized in that: The samples obtained in step S2 constitute a sample list, and the decision detection result described in step S3 or S4 is a hard decision detection result of selecting a sample with the smallest residual norm from the sample list, or a soft decision detection result of calculating a log-likelihood ratio from the sample list.
Citation Information
Patent Citations
Stochastic gradient geodesic Markov chain Monte Carlo method and apparatus
CN106599909A
Wet wood thermal parameter inversion method based on Bayesian reasoning
CN110489838A
MIMO-OFDM wireless signal detection method and system capable of channel matrix pre-processing during detection
US20220060365A1
