A mobile antenna enhanced integrated sensing communication transmit beamforming method

By using a unified antenna position mapping and product manifold optimization method, the problems of insufficient spatial degrees of freedom and optimization difficulties in the integrated sensing and communication system with movable antennas are solved. This method achieves joint optimization of antenna position and transmit beamforming, thereby improving the system's spectral efficiency and sensing and communication performance.

CN123055130APending Publication Date: 2026-11-06UNIV OF ELECTRONICS SCI & TECH OF CHINA +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611552203.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-09
Publication Date
2026-11-06

AI Technical Summary

Technical Problem

In existing integrated sensing and communication systems with movable antennas, the spatial degrees of freedom of fixed-position antenna arrays are insufficient, and the joint beamforming and position optimization of movable antennas suffer from strong coupling, complex constraints, and difficult solutions.

Method used

By employing a unified antenna position mapping and product manifold optimization method, a multi-channel input multi-channel output-inductive integrated MIMO-ISAC downlink system model is constructed. Combined with a parallel simplified quasi-Newton method, the transmit precoding matrix and array position vector are jointly optimized to achieve joint optimization of antenna position and transmit beamforming.

Benefits of technology

It effectively reduces beammap mean square error, improves spectral efficiency, and improves the trade-off between sensing and communication performance in integrated sensing and communication systems.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123055130A_ABST
    Figure CN123055130A_ABST
Patent Text Reader

Abstract

The application discloses a kind of mobile antenna enhanced integrated sensing communication transmit beamforming method, belong to wireless communication and sensing fusion technical field.The method includes: establishing mobile antenna enhanced integrated sensing communication downlink system model, and constructing the optimization target of considering multi-target beam pattern mean square error and multi-user spectral efficiency;Unified mapping is carried out to antenna position vector, and bounded movement and minimum spacing constraint are converted into unbound parameterization representation;Based on the mapped antenna position variable and transmit precoding matrix, construct product manifold, and adopt parallel simplified quasi-newton algorithm to carry out joint optimization, obtain transmit precoding matrix and mobile antenna position.The application can improve the compromise of sensing and communication performance of integrated sensing communication system, reduce beam pattern mean square error and improve spectral efficiency, and is suitable for multi-target integrated sensing communication scene.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of wireless communication and radar sensing fusion technology, specifically relating to an integrated sensing communication transmit beamforming method enhanced by a movable antenna, and more particularly to an integrated sensing communication transmit beamforming method based on unified antenna position mapping and product manifold optimization of a movable antenna. Background Technology

[0002] Integrated sensing communication achieves synergy between communication and sensing functions by sharing hardware platforms and spectrum resources, offering advantages in improving spectrum efficiency and energy efficiency. Transmit beamforming is a crucial means of realizing integrated sensing communication.

[0003] Existing integrated sensing and communication systems mostly employ fixed-position antenna arrays, whose array geometry is fixed and spatial degrees of freedom are limited, making it difficult to balance sensing and communication performance. Movable antenna arrays introduce additional spatial degrees of freedom by adjusting the antenna positions, providing a new means to improve the performance of integrated sensing and communication systems.

[0004] However, existing integrated sensing and communication methods using movable antennas typically employ alternating optimization, stepwise convex approximation, or projective gradient approaches, which suffer from strong variable coupling, slow convergence speed, sensitivity to initialization, and susceptibility to suboptimal solutions. Furthermore, the antenna position is subject to bounded movement and minimum spacing constraints, further increasing the optimization difficulty.

[0005] Therefore, there is an urgent need to provide a new integrated sensing communication transmit beamforming method with movable antenna enhancement to achieve joint optimization of antenna position and transmit beamforming. Summary of the Invention

[0006] The purpose of this invention is to provide an integrated sensing communication transmit beamforming method with movable antenna enhancement, in order to solve the problems of insufficient spatial degrees of freedom of fixed antennas in the prior art, as well as the problems of strong coupling, complex constraints, and difficulty in solving the joint beamforming and position optimization problem of movable antennas.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for integrated sensing communication transmit beamforming enhanced by a movable antenna includes the following steps:

[0009] S1: Establish a downlink system model of multi-channel input multi-channel output-inductive integrated MIMO-ISAC, and construct a joint optimization objective including beammap mean square error and spectral efficiency, as well as minimum spacing constraints, array length constraints, and total transmit power constraints;

[0010] S2: Perform a unified mapping on the array position vector of the movable antenna array, embedding the minimum spacing constraint and array length constraint into the unconstrained parameterized representation;

[0011] S3: Construct a product Riemannian manifold based on the mapped variables, and transform the original optimization problem into an unconstrained optimization problem on the manifold;

[0012] S4: A parallel simplified quasi-Newton method is used to jointly optimize the emission precoding matrix and the array position vector on the product manifold;

[0013] S5: Outputs the transmit precoding matrix and movable antenna position that satisfy the transmit power constraint.

[0014] Furthermore, S1 is specifically as follows:

[0015] In the MIMO-ISAC downlink system, the base station is equipped with a linear movable antenna array consisting of M antenna elements. The base station simultaneously provides services to U downlink users and detects T radar targets.

[0016] Let the array position vector of a one-dimensional movable antenna array be . ,in Let represent the position of the i-th antenna element, i = 1, 2, ..., M, and let . As a reference point, without loss of generality; and calculate the normalized guide vector of the one-dimensional movable antenna array;

[0017] Downlink ISAC Transmit Signal: During the downlink ISAC transmission interval, the base station uses a fixed array position vector throughout the entire time slot. and precoding matrix Perform downlink ISAC transmission; the transmission signal vector at the l-th symbol time is ,in The precoding matrix representing the overall ISAC baseband, and the synthesized symbol vector. Includes the former Communication symbol vector and the M radar symbol vectors ;

[0018] Downlink channel received signal: Obtain the downlink channel vector between the base station and the u-th user based on the normalized pilot vector. The received signal of the u-th user at the l-th symbol time It consists of the desired signal, the interference signal, and noise:

[0019]

[0020] Among them, superscript This indicates the conjugate transpose. Represents the precoding matrix The u-th element in the first U precoding vectors; This represents the u-th element among the first U composite symbol vectors, with the subscript j being the index that distinguishes it from the u-th element; This represents additive white Gaussian noise;

[0021] The total spectral efficiency of downlink communication is ,in The signal-to-interference ratio plus the noise ratio for the u-th user;

[0022] Transmit beam pattern specifications: angular sector Discretize into There are grid points, with an angle of [value]. ,in Determine the design resolution and direction The emission energy is:

[0023]

[0024] in, For direction The normalized guiding vector;

[0025] The mean square error of beam mode matching (MSE) is:

[0026]

[0027] in The actual scaling factor. For direction Beam weighting, For corresponding non-negative weights;

[0028] The combined movable antenna and beamforming design problem is

[0029]

[0030] in, To jointly optimize objectives, It is a scalar parameter; These are the minimum spacing constraint, array length constraint, and total transmit power constraint, respectively. This is the minimum distance between two adjacent antennas. This is the upper bound of the array length; Indicates the maximum available transmit power at the base station. This indicates taking the trace of the matrix.

[0031] Furthermore, S2 is specifically as follows:

[0032] First, the minimum spacing constraint The upper limit of the element spacing in the middle is ;in For real numbers, Array length constraint The corresponding lower bound is ;

[0033] Under the bounded minimum spacing constraint, each spacing in the array position vector is non-negative, and the sum of the spacings is bounded; a first auxiliary variable is introduced. And define the following scaling function as:

[0034]

[0035] in, This represents the scaling function for the k-th element in the first auxiliary variable. ; and The k-th element and the first auxiliary variable One element, This is the scaling control factor;

[0036] Introducing a second auxiliary variable and a Sigmoid function , used to control the total length of the movable antenna array;

[0037] Interval variable is defined as ;

[0038] The upper and lower bounds are derived from the interval variable. Given, represented as ;

[0039] The array position vector is mapped using a soft maximum function, specifically:

[0040]

[0041] The mapped antenna array spacing is:

[0042]

[0043] in The mapping interval function, based on the definition and upper and lower bounds of the interval variable, is:

[0044]

[0045] The total length of the movable antenna array is calculated as follows:

[0046]

[0047] also, .

[0048] Furthermore, S3 is specifically as follows:

[0049] By applying a unified array position vector mapping technique, the array position vector is rewritten as:

[0050]

[0051] in The reconstructed array position vector; the normalized guide vector of the movable antenna array is converted to... Then, the joint movable antenna and beamforming design problem was reformulated as follows:

[0052]

[0053] In the above formula, when When fixed, The closed-form solution is calculated as follows:

[0054]

[0055] The first matrix , For the reparameterized direction The emission energy;

[0056] The two auxiliary variables and the transport precoding matrix are represented as total variables. , , Tightened to ; and Placed On real Euclidean manifolds and one-dimensional real Euclidean manifolds Located on a complex spherical manifold, the product manifold PRM is constructed as follows:

[0057]

[0058] in Let each represent a product manifold. Real Euclidean manifolds, one-dimensional real Euclidean manifolds, and complex spherical manifolds;

[0059] By defining the tangent space of points on the product manifold, the Riemannian metric, and the transport operation between different tangent spaces, the problem is reformulated as a product manifold problem. The unconstrained optimization problem on the surface is denoted as: .

[0060] Furthermore, S4 is specifically as follows:

[0061] S4.1: First calculate exist Euclidean gradient Then, through orthogonal projection, the Euclidean gradient is... Projected onto the product manifold The Riemann gradient is obtained from ;

[0062] S4.2: Obtain the optimal descent direction by solving the Riemann-Newton equations. :

[0063]

[0064] in Representing Riemannianism; Parallel simplified quasi-Newton methods are utilized in The curvature information accumulated in a finite memory iteration constructs an approximate descent direction, and the parallel simplified quasi-Newton direction is defined as:

[0065]

[0066] This represents the inverse Riemannian approximation for each subvariable. Let x, z, and F be the directions of descent. , , The Riemann gradients of the three subvariables x, z, and F are given respectively; all inverse Riemann-Hessian submatrices are expressed using the same method as the total variables. The same limited memory update strategy, in At the nth update, the next inverse Riemann-Hessian approximation is constructed by calculating the following intermediate variables.

[0067]

[0068] in Represents the vector transfer operator. Step size, As the first intermediate variable, As the second intermediate variable, in the calculation Previously, the two intermediate variables were normalized to avoid simplifying the scaling of the quasi-Newton direction in parallel, resulting in the normalized first intermediate variable. and the normalized second intermediate variable ,Right now

[0069]

[0070] As a normalized intermediate variable representation, then, regarding The inverse Riemann-Hessian matrix is ​​updated to

[0071]

[0072] Where the updated variable The above expression has recursive properties; let's expand it. Next, among them Then, the inverse Riemann-Hessian matrix is ​​substituted into the parallel simplified quasi-Newton direction:

[0073]

[0074] Parallel simplification of quasi-Newton directions via Riemann gradient and intermediate variables The computation, during the iteration process, stores at most [number] units. Using a set of intermediate variables, the simplified quasi-Newton direction is calculated in parallel through a bidirectional recursive process.

[0075] S4.3: Use Update all variables intermediate variables Stored in an intermediate variable set according to a cautious update strategy.

[0076] Compared with existing technologies, this invention embeds antenna position constraints into a parameterized representation through a unified mapping, and combines product Riemannian manifold optimization to achieve joint design of antenna position and transmit beamforming. This can effectively reduce beam pattern mean square error, improve spectral efficiency, and improve the trade-off between sensing and communication performance in integrated sensing and communication systems. Attached Figure Description

[0077] Figure 1 A schematic diagram of an integrated sensing communication downlink system enhanced by a movable antenna;

[0078] Figure 2 The graph shows the changes in the objective function of the proposed method and the comparison method as a function of the number of iterations.

[0079] Figure 3 Comparison of transmit beam patterns under different methods;

[0080] Figure 4 A comparison chart of array layout and spectral efficiency under different methods;

[0081] Figure 5 The graph shows the mean square error of the beam pattern and the spectral efficiency as a function of the number of antennas.

[0082] Figure 6 A curve representing a trade-off between sensing and communication performance;

[0083] Figure 7 The graph shows the mean square error of the beam pattern and the spectral efficiency as a function of the length of the movable antenna array. Detailed Implementation

[0084] The present invention will be further described below with reference to the accompanying drawings. These embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. All equivalent substitutions or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0085] In this embodiment, a movable antenna-enhanced integrated sensing communication downlink system is considered, such as... Figure 1 As shown, the base station is equipped with a linear movable antenna array consisting of multiple antenna elements. By adjusting the antenna positions, additional spatial degrees of freedom are introduced, enabling the system to provide communication services to multiple downlink users while simultaneously sensing and detecting the target area. The system employs a time-division duplex operating mode, with uplink channel acquisition and echo reception occurring at different time intervals compared to downlink integrated sensing communication transmission.

[0086] Within a coherent time block, the antenna position, channel state information, and angle information remain unchanged. Based on the acquired channel information and prior sensing information, the base station designs the transmit precoding matrix and array position vector to achieve coordinated optimization of communication and sensing performance. Communication performance is measured by spectral efficiency, while sensing performance is characterized by the matching error between the transmit beammap and the desired beammap.

[0087] Based on the above system model, this embodiment provides an integrated sensing communication transmit beamforming method with movable antenna enhancement, including the following steps:

[0088] S1. Establish a MIMO-ISAC (Multi-channel input multi-channel output - inductive integration) downlink system model, and construct a joint optimization objective including beammap mean square error and spectral efficiency, as well as minimum spacing constraints, array length constraints, and total transmit power constraints;

[0089] In the MIMO-ISAC downlink system, the base station is equipped with a linear movable antenna array consisting of M antenna elements. The base station simultaneously provides services to U downlink users and detects T radar targets. The system operates in time-division duplex mode, with uplink channel acquisition and radar echo reception occurring at different time intervals compared to downlink integrated sensing communication transmission. Therefore, the base station (BS) transmits and receives at different times, and strong self-interference (SI) can be ignored in the baseband model, allowing for a focused design at the transmitter. In this coherent block, the MA (movable antenna) position, channel state information (CSI), and channel angle are assumed to be quasi-static.

[0090] Let the APV (array position vector) of a one-dimensional MA array be denoted as .

[0091]

[0092] in Let represent the position of the i-th antenna, i = 1, 2, ..., M. As a reference point, without loss of generality, the normalized guiding vector of a one-dimensional MA array is:

[0093]

[0094] in It is the carrier wavelength. It is a spatial angle. for .

[0095] To reduce mutual coupling and simplify array calibration, the spacing between any two adjacent antennas is subject to a minimum distance. The upper bound of the array length is limited. :

[0096]

[0097] in Representing the One constraint. To ensure the effectiveness of the far-field electromagnetic wave propagation model, select... , making The maximum length of the array is much smaller than the minimum user / target distance. The linear movable antenna array satisfies the minimum spacing constraint between adjacent elements and the total array length constraint; to satisfy the far-field array guidance model, the maximum array length is much smaller than the minimum user distance and the target distance.

[0098] Downlink ISAC Transmit Signal: Consider a block-fading MA-driven ISAC downlink. In each coherent module, the base station first acquires relevant user channel and angle information (such as AoD / AoA and target sector) through uplink training and radar echo processing, and then designs the downlink transmit signal based on these estimates. The training and receive processing phases are not explicitly modeled; instead, the generated CSI and angle parameters are considered known in the transmitter design problem. For a monobase single-BS ISAC, mitigating strong self-interference (SI) leakage from the transmitter to the receiver chain is itself a key design challenge, typically achieved through the SI channel matrix (e.g., This study focuses on explicit modeling using analog / digital cancellation or SI-sensing beamforming. The framework is directly applicable to TDD (Time Division Duplex) based ISAC systems or architectures with separate transmit and receive apertures, where downlink beamforming can be designed solely based on CSI and angle information. Explicit SI modeling is then required (e.g., in systems with non-negligible CSI). In a full-duplex (FD) single-base deployment, the launch design developed here can be embedded into the SI sensing processing chain.

[0099] During the downlink ISAC transmit interval, the BS uses a fixed MA position vector throughout the entire time slot. and precoding matrix Perform downlink ISAC transmission. The transmitted signal vector for the l-th symbol time is...

[0100]

[0101] in This represents the overall ISAC baseband precoding matrix, divided by column into... .forward precoded vectors It is used for both downlink communication and radar sensing, while the remaining M precoded vectors It is specifically used for radar sensing. Synthetic symbol vector. Includes the former Downlink User Communication Symbol Vector and radar symbols , Let represent the synthesized symbol vector of the i-th precoding vector. Assume that the communication symbols and radar symbols are uncorrelated, and normalize them so that... , in Indicates expectation, superscript This indicates the conjugate transpose. Represents the identity matrix.

[0102] Total transmit power is limited by

[0103]

[0104] in This indicates the maximum available transmit power at BS. This indicates taking the trace of the matrix.

[0105] Furthermore, communication performance is characterized by downlink signal-to-interference-plus-noise ratio (SIR) and spectral efficiency, specifically:

[0106] 2) The communication metric adopts a geometric multipath channel model under the far-field assumption. For the first... The field response vector of the q-th user and the q-th path is guided by the normalized vector. Given, among which This represents the starting angle (AoD) of the path. Assume there is... There are several propagation paths, within the considered groups, between the BS and the user. The downlink channel vector between them is modeled as

[0107]

[0108] in It is the complex gain of the q-th path. A MA-specific channel estimation scheme with reasonable training overhead is provided, where downlink communication performance is characterized by user signal-to-interference-plus-noise ratio and spectral efficiency. In the block-by-block design, These are considered to be known parameters obtained during previous uplink training.

[0109] Given downlink channel vector ,user The received signal is

[0110]

[0111] The first two terms on the left side of the equation are the desired signal and the interference signal, respectively. This represents additive white Gaussian noise. This represents the noise variance. The signal-to-interference-plus-noise ratio (SINR) for user u is...

[0112]

[0113] The total spectral efficiency (SE) of downlink communication is

[0114]

[0115] Sensing performance is characterized by the mean square error of beammap matching, including:

[0116] 3) Sensing metrics are evaluated using classic transmit beam pattern metrics: Within the downlink ISAC transmit interval, the same signal serving the user will also illuminate the sensing scene. The desired angular power pattern is specified within the discretized sector, and the MA position and pre-encoder are optimized to make the actual beam pattern approximate the target shape, thereby controlling the echo signal-to-noise ratio distribution and angle estimation accuracy.

[0117] angular sector Discretize into There are grid points, with an angle of [value]. ,in Determine the design resolution (e.g., for a 1◦ grid, Based on prior information (e.g., angle of incidence (AoA) estimates for the first few blocks), Divided into:

[0118] target area The location corresponding to the desired target; for Beam weight Set it close to 1 to obtain a higher gain.

[0119] Interference / clutter area This corresponds to a known interfering entity or dominant clutter; for Beam weight Set to small (or approximately 0) to achieve a deep indentation. The background area is... No strict shaping is required. Take the middle value.

[0120] Then, a non-negative weight is assigned to each grid point. Main lobe The larger one corresponding to the range Emphasizing the accuracy of the main lobe, Larger Enhanced low sidelobe, smaller middle This provides more freedom for communication. Therefore, for This indicates different perceptual priorities, among which It includes beam weights for all angles.

[0121] Far-field array to direction The response is determined by direction. The corresponding normalized guiding vector Given. In covariance normalization Down, direction The emission energy is

[0122]

[0123] The mean square error of beam mode matching (MSE) is

[0124]

[0125] in This is the actual scaling factor. (Scalar) Decouple the desired angular shape from its absolute power: The relative distribution at a specified angle, while The overall level will be adjusted to [the following] and A determined actual transmit power is used to avoid pure scale mismatch dominating MSE. Beam pattern-MSE is a standard for phased array and MIMO radar waveform design and is widely used as a substitute indicator for sensing quality. Suppressing the allocation of more power during the process can improve the echo signal-to-noise ratio in the target direction and tighten classical limits such as CRB, thereby reducing the root mean square error (RMSE) of the direction of arrival (DOA) estimation.

[0126] From the perspective of multi-objective optimization, sensing and communication objectives are fundamentally different: beam pattern MSE characterizes the desired angular power distribution. The approximation of the target is determined by the approximation of the target value, while SE measures the downlink throughput transmitted to users. To explore the Pareto trade-off between these two objectives, standard weighting and scalarization are employed, where the scalar parameter... Control the relative priority between sensing and communication.

[0127] For a given reference frame, the joint MA and beamforming design problem is as follows:

[0128]

[0129] although It is convex, but due to the MSE and SE terms... and Coupling dependency, target Highly nonconvex. Existing MA-ISAC designs typically rely on AO (Alternating Optimization), SCA (Successive Convex Approximation), and SPGA (Simplified Spatial Generalization), which can lead to slow convergence and high initialization sensitivity. This optimization problem is solved by developing a UM–PM (Unified Mapping-Integral Manifold) framework, embedding MA constraints into variable parameterization, thus achieving… and The overall joint optimization is achieved without the need for explicit alternating decomposition.

[0130] S2. Perform a unified mapping on the array position vector of the movable antenna array, embedding the minimum spacing constraint and array length constraint into the unconstrained parameterized representation;

[0131] Despite the BMS (Bounded Minimum Spacing) constraint on APV While the constraints are convex, enforcing them during optimization remains challenging. Existing MA-ISAC work typically handles these constraints through SCA-based linearization or through projective feasibility recovery in gradient-based SPGA methods. The former introduces repetitive approximations, while the latter is affected by the sequential nature of minimum spacing constraints, often leading to inaccurate corrections and convergence distortions. To avoid iterative linearization and post-update predictions, a unified APV mapping was developed, embedding BMS constraints into a smooth, unconstrained parameterization. This eliminates explicit constraint handling, enabling seamless joint optimization.

[0132] Specifically, the spacing variable is constructed using auxiliary variables. Firstly, since the total length of the MA array is bounded, The upper limit of the element spacing in the middle is

[0133]

[0134] Among real numbers The above formula represents the maximum allowable value for the spacing between a single segment. The upper bound indicates the constraint... The maximum segment spacing achievable when the total available array length is fully utilized. The above inequality is not an additional constraint, but a natural limit determined by the BMS conditions. Similarly, since each spacing is a lower bound, The lower bound of the total array length in the array is .

[0135]

[0136] Under BMS constraints, each spacing in the APV is non-negative, and the sum of the spacings is bounded. A first auxiliary variable is introduced. And define the following scaling function as:

[0137]

[0138] in, This represents the scaling function for the k-th element in the first auxiliary variable. ; and The k-th element and the first auxiliary variable One element, This is the scaling control factor;

[0139] Therefore, when hour, and .when At this point, controlling the sum of the spacing becomes difficult, which limits the depth of field for adjusting the length of the MA array. Therefore, a second auxiliary variable is introduced. and a Sigmoid function , used to control the total length of the MA array.

[0140] Finally, the interval variable is defined as

[0141]

[0142] Clearly, it can be deduced that the upper and lower bounds are determined by the interval variable. Given, represented as

[0143]

[0144] The array position vector is mapped using a soft maximum function, specifically:

[0145] Using the spacing variable, APV can be mapped to a soft maximum sum function of (x, z), as follows:

[0146]

[0147] By mapping APV to the above equation, the BMS constraint in the soft maximum sum expression is expressed through the scaling function. and Inherent satisfaction, while parameters Remaining unconstrained. In other words, the proposed unified APV mapping technique transforms APV into an unconstrained representation in real Euclidean space, with BMS constraints fully embedded in the equivalent form.

[0148] First, the spacing between the mapped antenna arrays is

[0149]

[0150] in This represents the mapping interval function. Based on the definition and upper and lower bounds of the interval variable, we can obtain...

[0151]

[0152] in Contains This corresponds to the lower bound of the above formula. Furthermore, when certain exist and Time (all others) ),constraint The remaining array lengths are concentrated in the first... In this section, the upper bound of the above formula is given.

[0153] Then, the total length of the mapping array of the antenna array is calculated as follows:

[0154]

[0155] In addition, due to It is obvious We can conclude that... .

[0156] S3. Construct a product Riemannian manifold based on the mapped variables, and transform the original optimization problem into an unconstrained optimization problem on the manifold;

[0157] This step developed the UM-PM method to solve the problem in three steps. First, a unified APV mapping technique was applied to express BMS constraints (i.e., and The embedded parameterization process eliminates the need to explicitly enforce them in joint multimolecule and beamforming design problems. Then, the RPM (integrated Riemannian manifold) is constructed as... The feasible domain is determined, thus reformulating the problem as an unconstrained optimization task on RPM. Finally, a parallel simplified quasi-Newton (PSQN) iterative method for manifold structures is developed to achieve... Synchronous updates. Details are as follows:

[0158] First, by applying unified APV mapping technology, APV is rewritten as...

[0159]

[0160] in This represents the reconstructed array position vector. Furthermore, the normalized guide vector of the MA array can be converted to...

[0161]

[0162] Then, the problem was rephrased as

[0163]

[0164] In the above formula, when When fixed, The closed-form solution is calculated as follows:

[0165]

[0166] The first matrix , Direction after reparameterization The emission energy. However, due to Non-convex, and variable The coupling between them is strong, and solving the problems after restatement remains challenging. Therefore, the RPM space is constructed as follows: This allows us to re-formulate the problem as an unconstrained optimization problem by opening up the solution space.

[0167] First, the two auxiliary variables and the transmission precoding matrix are represented as total variables, i.e. . and It can be placed naturally It is situated on a one-dimensional real Euclidean manifold (REM). Furthermore, in order to fully utilize available transmit power and improve ISAC performance, Tightened to Under this constraint, Located on a complex spherical manifold (CSM), which inherently satisfies the power normalization condition, the PRM is constructed as follows:

[0168]

[0169] in Denotes the product manifold PRM. 3D REM, 1D REM, and CSM. Specifically, they are written as:

[0170]

[0171] The product manifold includes the tangent space, Riemannian metric, and vector transfer operator, specifically:

[0172] In order to optimize It is necessary to define the tangent space of the PRM point, the Riemann metric, and the transport operations between different tangent spaces, as follows:

[0173] Cut space: PRM exist The tangent space at a point is the Cartesian product of the tangent spaces of the submanifolds:

[0174]

[0175] in, , , , Indicates in , , , The tangent space at the point,

[0176]

[0177]

[0178]

[0179] in , , These are the corresponding tangent vectors. This indicates taking the real part.

[0180] Riemannian metric: The Riemannian metric on the product space is defined as the standard inner product on the product space, which is considered to be... For Euclidean space The Riemannian manifold. For any of The tangent vector on the tangent space corresponding to another point on the product manifold that distinguishes it from V. The Riemann metric is defined as

[0181]

[0182] in, To distinguish it from the tangent vector of the submanifold at another point of V, It is the dot product of vectors;

[0183] Vector Transfer: To ensure that the search direction remains unchanged across multiple iterations, a vector transfer operator is defined. It is represented as the point Tangent vectors on the corresponding tangent space Mapped to another point In the corresponding tangent space, it is as follows:

[0184]

[0185] The Euclidean fraction remains unchanged, while the component on the complex spherical manifold is projected onto the new tangent space.

[0186] Based on all the above factors, the problem is rephrased as follows: The unconstrained optimization problem on the surface is denoted as:

[0187]

[0188] S4. The transmit precoding matrix and antenna position variables are jointly optimized on the product manifold using the parallel simplified quasi-Newton (PSQN) method;

[0189] because The periodic nonconvexity and The strong coupling between the two makes first-order methods such as Steepest Descent (SD) and Conjugate Gradient Descent (CGD) prone to being trapped in local optima. To alleviate this problem, a second-order optimization strategy is adopted, utilizing curvature information through the proposed PSQN method. Specifically, PSQN approximates the inverse Hessian matrix by using intermediate variables stored in a finite number of past iterations, thereby significantly reducing computational complexity. The PSQN method consists of the following three main steps:

[0190] S4.1: Steps for calculating the Riemann gradient;

[0191] First calculate exist Euclidean gradient

[0192]

[0193] in To find the Euclidean gradient operator, This represents taking the conjugate of the complex number, and the specific gradient is calculated as follows:

[0194]

[0195] Then, by using the Euclidean gradient Projected into PRM space The Riemann gradient is obtained from the orthogonal projection. According to the defined orthogonal projection, the Riemann gradient can be obtained as follows:

[0196]

[0197] The corresponding Riemann gradient value is equal to the Euclidean gradient.

[0198] S4.2: The search direction is obtained by calculating the approximate inverse Riemann-Hessian matrix;

[0199] The optimal descent direction is obtained by solving the Riemann-Newton equations. :

[0200]

[0201] in Representing Riemann and Hessian, This represents the Riemann gradient. To avoid the high computational cost and potential uncertainty of the exact Hessian matrix, the PSQN method utilizes... The curvature information accumulated in a finite memory iteration constructs an approximate descent direction.

[0202] PSQN direction is defined as

[0203]

[0204] This represents the inverse Riemannian approximation for each subvariable. Let x, z, and F be the directions of descent. , , Let x, z, and F be the Riemann gradients of the three subvariables. All inverse Riemann-Hessian submatrices are expressed using the same method as the total variables. The same limited memory update strategy. Specifically, in At the nth update, the next inverse Riemann-Hessian approximation is constructed by calculating the following intermediate variables. :

[0205]

[0206] in Represents the vector transfer operator. Step size, As the first intermediate variable, This is the second intermediate variable. In the calculation... Previously, intermediate variables were normalized to avoid scaling in the PSQN direction, i.e.

[0207]

[0208] As a normalized intermediate variable representation, then, regarding The inverse Riemann-Hessian matrix is ​​updated to

[0209]

[0210] Where the updated variable The above expression has recursive properties; let's expand it. Next, among them Then, substituting it into the PSQN direction definition, the PSQN direction definition is rewritten as follows:

[0211]

[0212] The PSQN direction can be determined via the Riemann gradient. and intermediate variables Computation. To reduce storage and computational complexity, the maximum storage during iteration is [number missing]. A set of intermediate variables. Length is... The storage space is defined as:

[0213]

[0214] in express The Each element. Then, using the stored intermediate variables... The PSQN direction is calculated through a bidirectional recursive process.

[0215] S4.3: Optimize variable updates;

[0216] Update feasible variables And store intermediate variables First, use the PSQN direction. It can be updated. for

[0217]

[0218] Among them, Represents a product manifold variable backtracking operation, step size The result is obtained using the Armijo Linear Search (ALS) strategy. The ASL strategy finds the smallest non-negative integer. To satisfy monotony

[0219]

[0220] in .

[0221] Then, the intermediate variables are stored according to the Cautious Update Criterion (CUF) strategy. This strategy ensures that the inverse Hessian approximation remains positive definite, accepting only curvature information that satisfies the standard BFGS condition. In the nth iteration, the threshold calculation performed by CUF is:

[0222]

[0223] According to the CUF strategy exist Storage within the stack follows a left-to-right write-to-stack principle. All existing... exist All must remain in the corresponding previous iteration In the tangent space. Assume If m sets of intermediate variables are currently stored, the following situation may occur:

[0224] if : Dissatisfied. The vector transfer operation was executed, and the latest intermediate variables are stored as follows:

[0225]

[0226] in express The One element;

[0227] if : It is full. The leftmost intermediate variable is discarded, and the vector transition operation is applied to the remaining variables. The latest intermediate variable is stored as follows:

[0228]

[0229] If the threshold conditions for the cautious update criterion are not met, it is still necessary to... Transform existing intermediate variables on into corresponding The tangent space, i.e. .

[0230] S5. Output the transmit precoding matrix and the position of the movable antenna that satisfy the transmit power constraint.

[0231] Simulation results are presented to evaluate the performance of the proposed UM–PM method on MA arrays, denoted as UM–PM(MA). For comparison, consider the following two benchmark schemes:

[0232] CGD (FPA): ULA spacing is Furthermore, the FPA of F is optimized using the CGD method.

[0233] AO–SPGA (MA): An AO frame with an MA array. For fixed... , Updated using the CGD method; for fixed... , The projection is updated to the BMS feasible set via the SPGA step.

[0234] Unless otherwise specified, the system parameter settings are as follows. For MA arrays, the minimum element spacing is set to... The maximum aperture is .in It is relative to the reference. scaling factor , Let be the aperture length of a uniform linear array. For communication tasks, consider the downlink. User, noise power And each user has Path gain They are independently and identically distributed, and follow a constant distribution. The user's AoA is plotted at approximately Within a small sector, so that the communication area is separated from the sensing main lobe. For the sensing task, the beam patterns and weights corresponding to the target, interference, and background angles are respectively determined by... , , and , To specify, the specific choice depends on:

[0235] Without loss of generality, consider the common case of non-interference multi-beams, where , , ;set up , , and .

[0236] In the case of multi-beam interference, the target sector remains the same as before, that is And add interference sectors ,in The desired beam pattern is set to... ,express ,otherwise The weight is then selected when hour ,when hour This is to emphasize the suppression of sidelobes in the interference sector.

[0237] Figure 2 The UMPM (MA) and AO–SPGA (MA) were demonstrated in and At that time, weighted cost exist The convergence behavior is examined. For both array sizes, UM–PM (MA) reduces the objective value faster and converges to significantly lower values, while AO–SPGA (MA) converges more slowly and saturates at higher cost levels. The AO–SPGA (MA) curve is not strictly monotonic, consistent with its design: each p-update applies a fixed-step gradient followed by a global projection of the BMS feasible set within the block coordinate AO ring. This predicted GA step does not guarantee a reduction in global cost for each iteration, leading to… Figure 3 Small oscillations were observed. Figure 4 The proposed UMPM (MA) method was compared with the CGD (FPA) and AO–SPGA (MA) baselines. Down and . Figure 3 The convergence curves in (a) and (d) of Figure 3 show that UMPM achieves the lowest weighted cost and the optimal MSE–SE tradeoff, while AO–SPGA (MA) lies between UM–PM (MA) and CGD (FPA). At that time, UM–PM (MA) converged to a sensing MSE of −2.24 dB and an SE of 5.82 bps / Hz, compared to CGD (FPA) at 1.32 dB and 4.11 bps / Hz, and AO–SPGA (MA) at −0.29 dB and 5.56 bps / Hz. For The final MSE / SE pairs were (UM–PM (MA) -4.75 dB, 6.89 bps / Hz, CGD (FPA) -2.99 dB, 6.18 bps / Hz) and AO–SPGA (MA) -4.53 dB, 6.64 bps / Hz, confirming that the proposed framework can always achieve the optimal operating point.

[0238] Figure 3 (b) and Figure 3 The beam pattern in (e) further illustrates this behavior. Compared to FPA, both MA-based schemes significantly improve the main lobe flatness and side lobe suppression within the target sector, with the UM–PM(MA) scheme achieving the most uniform power distribution and the deepest concavity within the sector. Figure 3 The array layout corresponding to (c) shows that all MA schemes deviate from the unified ULA structure while satisfying the BMS constraints. Figure 3 (f) in the middle shows and Time SE performance and user signal-to-noise ratio. Throughout the entire signal-to-noise ratio (SNR) range, UM–PM(MA) consistently achieves the lowest standard error (SE), while AO–SPGA(MA) shows a significant advantage over CGD(FPA). The performance improvement of these two moving average (MA) based schemes compared to a fixed FPA benchmark is roughly equivalent.

[0239] To assess scalability, Figure 4 Drawing in the sensor guidance ( ) and communication guidance ( Under the settings, the sensing MSE and SE are related to the number of antennas. The relationship. In a sensor-oriented context ( Both probabilities benefit from a larger M, but UM–PM (MA) always achieves a significantly lower MSE.

[0240]

[0241] The additional spatial depth of field provided by MA repositioning allows for finer beam pattern control than fixed ULA, especially in... When it increases. In the case of communication ( The advantages of MA become even more apparent. With... With the increase in UM-PM, the SE continues to grow, while the SE of FPA-based systems will eventually degrade or even collapse. Although the multi-user interference term remains roughly constant, the sensing interference term increases with M. For data that can only be transmitted through M... With FPA acting on a fixed geometry, the available DoF is insufficient to suppress this increasing disturbance, leading to degradation of SINR and SE. In contrast, UM–PM joint adaptation... and Reshape the channel itself, and in Maintain SE gain when increasing.

[0242] Figure 5 Demonstrates the use of sweep weights The Pareto boundaries of UM–PM (MA) and CGD (FPA) are obtained. The ideal region is the lower right corner, with low MSE but high SE. For and The boundary of UM–PM (entity) always dominates CGD (FPA) (dashed line): for any target MSE, UM–PM achieves a higher SE, while for any target SE, it produces a lower MSE. For example, in At the time, the representative operating point at the UM–PM boundary showed an SE increase of approximately 0.65 bps / Hz and an MSE decrease of approximately 2.32 dB compared to the baseline. Furthermore, when When the value is increased from 8 to 12, the Pareto boundary of UM–PM moves more significantly toward the ideal angle than that of FPA.

[0243] Figure 6 The maximum physical aperture of the MA array (by scaling factor) was investigated. (Control) in and The effect of time. The FPA baseline remains unchanged as a static reference. In contrast, UM–PM (MA) is significantly dependent on the available aperture. As LD increases from 1.3 to approximately 1.9, MSE improves significantly due to higher angular resolution and lower sidelobes. MSE tends to saturate, while SE continues to grow, reflecting that the optimizer is beginning to trade a small loss of sensing accuracy for a significant increase in throughput. Quantitatively, in At that time, the MSE of UM–PM was about 3.15 dB lower than the FPA baseline, and the SE was 0.35 bps / Hz higher.

[0244] Figure 7 The proposed UM–PM (MA) method was compared with... and CGD (FPA) and AO–SPGA (MA) baselines under multi-beam interference conditions. Figure 7 The convergence curves in (a) and (d) of 7 show that even though the sensing weights of UM–PM (MA) and AOSPGA (MA) are larger ( vs CGD (FPA) The proposed UM-PM (MA) still achieves the lowest weighted cost and the best MSE-SE tradeoff, while AO-SPGA (MA) falls between UM-PM (MA) and CGD (FPA), and because it predicts SPGA in... Visible oscillations appear due to updates on the surface. At that time, UM–PM (MA) converged to a sensing MSE of −5.25 dB and an SE of 9.139 bps / Hz, while CGD (FPA) was −2.04 dB and 5.81 bps / Hz, and AO–SPGA (MA) was −3.75 dB and 9.129 bps / Hz. For The final MSE / SE pairs are UMPM (MA) (−7.04 dB, 8.47 bps / Hz), CGD (FPA) (−3.70 dB, 6.41 bps / Hz), and AO–SPGA (MA) (−5.91 dB, 8.50 bps / Hz), which confirms that the proposed framework can achieve the lowest sensing error in the presence of strong interference while maintaining competitive SE.

[0245] Figure 7 Figures (b) in 7 and (e) in 7 illustrate the resulting beam patterns. Both MA-based schemes form deep grooves around the interference region while maintaining well-shaped main lobes in both target regions, while CGD (FPA) exhibits higher residual side lobes near the interference layer. Compared to AOSPGA (MA), UM–PM (MA) has slightly cleaner grooves and more uniform power in the target sector, which explains its lower MSE. Figure 7 The array layout in (c) further illustrates that the MA-based scheme adjusts the element positions from outside the uniform ULA while satisfying the BMS constraint, while CGD (FPA) maintains a fixed geometry. Figure 7 Figure (f) in the figure plots the relationship between SE and user signal-to-noise ratio, showing that for and Both MA-based schemes consistently outperform the FPA baseline across the entire signal-to-noise ratio range, with UMPM (MA) and AO–SPGA (MA) achieving nearly overlapping SE curves.

Claims

1. A method for integrated sensing communication transmit beamforming enhanced by a movable antenna, characterized in that, Includes the following steps: S1: Establish a downlink system model of multi-channel input multi-channel output-inductive integrated MIMO-ISAC, and construct a joint optimization objective including beammap mean square error and spectral efficiency, as well as minimum spacing constraints, array length constraints, and total transmit power constraints; S2: Perform a unified mapping on the array position vector of the movable antenna array, embedding the minimum spacing constraint and array length constraint into the unconstrained parameterized representation; S3: Construct a product Riemannian manifold based on the mapped variables, and transform the original optimization problem into an unconstrained optimization problem on the manifold; S4: A parallel simplified quasi-Newton method is used to jointly optimize the emission precoding matrix and the array position vector on the product manifold; S5: Outputs the transmit precoding matrix and movable antenna position that satisfy the transmit power constraint.

2. The integrated sensing communication transmit beamforming method with movable antenna enhancement according to claim 1, characterized in that, S1 is specifically as follows: In the MIMO-ISAC downlink system, the base station is equipped with a linear movable antenna array consisting of M antenna elements. The base station simultaneously provides services to U downlink users and detects T radar targets. Let the array position vector of a one-dimensional movable antenna array be . ,in Let represent the position of the i-th antenna element, i = 1, 2, ..., M, and let . As a reference point, without loss of generality; and calculate the normalized guide vector of the one-dimensional movable antenna array; Downlink ISAC Transmit Signal: During the downlink ISAC transmission interval, the base station uses a fixed array position vector throughout the entire time slot. and precoding matrix Perform downlink ISAC transmission; the transmission signal vector at the l-th symbol time is ,in The precoding matrix representing the overall ISAC baseband, and the synthesized symbol vector. Includes the former Communication symbol vector and the M radar symbol vectors ; Downlink channel received signal: Obtain the downlink channel vector between the base station and the u-th user based on the normalized pilot vector. The received signal of the u-th user at the l-th symbol time It consists of the desired signal, the interference signal, and noise: Among them, superscript This indicates the conjugate transpose. Represents the precoding matrix The u-th element in the first U precoding vectors; This represents the u-th element among the first U composite symbol vectors, with the subscript j being the index that distinguishes it from the u-th element; This represents additive white Gaussian noise; The total spectral efficiency of downlink communication is ,in The signal-to-interference ratio plus the noise ratio for the u-th user; Transmit beam pattern specifications: angular sector Discretize into There are grid points, with an angle of [value]. ,in Determine the design resolution and direction The emission energy is: in, For direction The normalized guiding vector; The mean square error of beam mode matching (MSE) is: in The actual scaling factor. For direction Beam weights, For corresponding non-negative weights; The design problem of combined movable antenna and beamforming is in, To jointly optimize objectives, It is a scalar parameter; These are the minimum spacing constraint, array length constraint, and total transmit power constraint, respectively. This is the minimum distance between two adjacent antennas. This is the upper bound of the array length; Indicates the maximum available transmit power at the base station. This indicates taking the trace of the matrix.

3. The integrated sensing communication transmit beamforming method with movable antenna enhancement according to claim 2, characterized in that, S2 is specifically as follows: First, the minimum spacing constraint The upper limit of the element spacing in the middle is ;in For real numbers, Array length constraint The corresponding lower bound is ; Under the bounded minimum spacing constraint, each spacing in the array position vector is non-negative, and the sum of the spacings is bounded; a first auxiliary variable is introduced. The following scaling function is defined as follows: in, This represents the scaling function for the k-th element in the first auxiliary variable. ; and The k-th element and the first auxiliary variable One element, This is the scaling control factor; Introducing a second auxiliary variable and a Sigmoid function , used to control the total length of the movable antenna array; Interval variable is defined as ; The upper and lower bounds are derived from the interval variable. Given, represented as ; The array position vector is mapped using a soft maximum function, specifically: The mapped antenna array spacing is: in The mapping interval function, based on the definition and upper and lower bounds of the interval variable, is: The total length of the movable antenna array is calculated as follows: also, .

4. The integrated sensing communication transmit beamforming method with movable antenna enhancement according to claim 3, characterized in that, S3 is specifically as follows: By applying a unified array position vector mapping technique, the array position vector is rewritten as: in The reconstructed array position vector; the normalized guide vector of the movable antenna array is converted to... Then, the joint movable antenna and beamforming design problem was reformulated as follows: In the above formula, when When fixed, The closed-form solution is calculated as follows: The first matrix , For the reparameterized direction The emission energy; The two auxiliary variables and the transport precoding matrix are represented as total variables. , , Tightened to ; and Placed On real Euclidean manifolds and one-dimensional real Euclidean manifolds Located on a complex spherical manifold, the product manifold PRM is constructed as follows: in Let each represent a product manifold. Real Euclidean manifolds, one-dimensional real Euclidean manifolds, and complex spherical manifolds; By defining the tangent space of points on the product manifold, the Riemannian metric, and the transport operation between different tangent spaces, the problem is reformulated as a product manifold problem. The unconstrained optimization problem on the surface is denoted as: .

5. The integrated sensing communication transmit beamforming method with movable antenna enhancement according to claim 4, characterized in that, S4 is specifically as follows: S4.1: First calculate exist Euclidean gradient Then, through orthogonal projection, the Euclidean gradient is... Projected onto the product manifold The Riemann gradient is obtained from ; S4.2: Obtain the optimal descent direction by solving the Riemann-Newton equations. : in Representing Riemannianism; Parallel simplified quasi-Newton methods are utilized in The curvature information accumulated in a finite memory iteration constructs an approximate descent direction, and the parallel simplified quasi-Newton direction is defined as: This represents the inverse Riemannian approximation for each subvariable. Let x, z, and F be the directions of descent. , , The Riemann gradients of the three subvariables x, z, and F are given respectively; all inverse Riemann-Hessian submatrices are expressed using the same method as the total variables. The same limited memory update strategy, in At the nth update, the next inverse Riemann-Hessian approximation is constructed by calculating the following intermediate variables. in Represents the vector transfer operator. Step size, As the first intermediate variable, As the second intermediate variable, in the calculation Previously, the two intermediate variables were normalized to avoid simplifying the scaling of the quasi-Newton direction in parallel, resulting in the first normalized intermediate variable. and the normalized second intermediate variable ,Right now As a normalized intermediate variable representation, then, regarding The inverse Riemann-Hessian matrix is ​​updated to Where the updated variable The above expression has recursive properties; let's expand it. Next, among them Then, the inverse Riemann-Hessian matrix is ​​substituted into the parallel simplified quasi-Newton direction: Parallel Simplification of Quasi-Newton Directions via Riemann Gradient and intermediate variables The computation, during the iteration process, stores at most [number] units. Using a set of intermediate variables, the simplified quasi-Newton direction is calculated in parallel through a bidirectional recursive process. S4.3: Use Update all variables intermediate variables Stored in an intermediate variable set according to a cautious update strategy.

6. The integrated sensing communication transmit beamforming method with movable antenna enhancement according to claim 5, characterized in that, The normalized guide vector of the one-dimensional movable antenna array is calculated as follows: in It is the carrier wavelength. It is a spatial angle. for .

7. The integrated sensing communication transmit beamforming method with movable antenna enhancement according to claim 6, characterized in that, The definition of the tangent space of points on the product manifold, the Riemannian metric, and the transport operation between different tangent spaces are as follows: Tangent space: Product manifold exist The tangent space at a point is the Cartesian product of the tangent spaces of the submanifolds: in, , , , Indicates in , , , The tangent space at the point, , , ;in , , These are the tangent vectors of the corresponding submanifolds. Indicates taking the real part; Riemannian metric: product manifold The Riemannian metric on the product space is defined as the standard inner product on the product space, which is considered to be... For Euclidean space Riemannian manifolds, for tangent vector in The tangent vector on the tangent space corresponding to another point on the product manifold that is distinct from V. The Riemann metric is defined as in, To distinguish it from the tangent vector of the submanifold corresponding to another point of V, It is the dot product of vectors; Vector Transfer: To ensure that the search direction remains unchanged across multiple iterations, a vector transfer operator is defined. It is represented as the point Tangent vectors on the corresponding tangent space Mapped to another point In the corresponding tangent space, it is as follows: The Euclidean fraction remains unchanged, while the component on the complex spherical manifold is projected onto the new tangent space.

8. The integrated sensing communication transmit beamforming method with movable antenna enhancement according to claim 7, characterized in that, The Riemann gradient is calculated in detail as follows: First, calculate the Euclidean gradient. : in To find the Euclidean gradient operator, This indicates taking the conjugate of the complex number, and the specific gradient is calculated as follows: Based on orthogonal projection, the Riemann gradient is obtained as follows: The corresponding Riemann gradient value is equal to the Euclidean gradient.

9. The integrated sensing communication transmit beamforming method with movable antenna enhancement according to claim 8, characterized in that, S4.3 is as follows: Length is The storage space is defined as ; Use the descent direction ,renew for in Represents a product manifold variable backtracking operation, step size The intermediate variables are obtained using the Armijo linear search strategy; then, they are stored according to a cautious update criterion strategy. ,exist In the nth iteration, the threshold for implementing the cautious update criterion is calculated as follows: exist The storage follows a left-to-right stack write principle, with all existing data being written to the stack. exist All must remain in the corresponding previous iteration In the tangent space; assuming If m sets of intermediate variables are currently stored in the current database, the following situation may occur: if : Dissatisfied, the vector transfer operation was executed, and the latest intermediate variables are stored as follows: in express The One element; if : It is full. The leftmost intermediate variable is discarded, and the vector transition operation is applied to the remaining variables. The latest intermediate variable is stored as follows: If the threshold conditions for the cautious update criterion are not met, it is still necessary to... Transform existing intermediate variables on into corresponding The tangent space, i.e. .