12 results about "Phase quantization" patented technology

Various aspects of the present disclosure generally relate to wireless communication. In some aspects, a device may perform, using vector data, an amplitude quantization based at least in part on an initial amplitude observation. The device may calculate one or more phase quantization bits based at least in part on the amplitude quantization. The device may determine a first bit allocation for amplitude and phase, wherein the first bit allocation is an initial amplitude and phase bits allocation for different elements of the vector data. The device may iteratively modify the first bit allocation, to obtain a second bit allocation, based at least in part on an amplitude bits adjustment and quantization, wherein the second bit allocation is a final amplitude and phase bits allocation for the different elements of the vector data. Numerous other aspects are described.

Methods, systems, and devices for wireless communications are described. A user equipment (UE) may report quantized information using a threshold payload utilization. For example, the UE may perform precoding matrix indicator (PMI) or channel reporting quantization to fill a total overhead allocated for quantization. The UE may include, in a channel state information (CSI) report, sets of amplitude quantization bits and phase quantization bits for one or more PMIs, the channel reporting information, or both. A quantity of the amplitude quantization bits may be based on a quantity of non-zero amplitude coefficients for reporting and a quantity of quantization bits associated with amplitude quantization. A quantity of the phase quantization bits may be based on the total overhead and the quantity of the amplitude quantization bits (e.g., to fill the remaining resources of the allocated overhead with the phase quantization bits).The UE may transmit the resulting CSI report.

The invention discloses a 1553B bus decoder design method for quickly locking asynchronous data based on phase quantization, and belongs to the technical field of information processing. According to the method, a sampling clock signal with fixed cycle time is quantized and then stored in an LUT table, when asynchronous random data comes, a decoder obtains the edge of the asynchronous random signal, meanwhile, the front-back phase relation and the phase error of the rising edge of an edge signal and the falling edge of the sampling clock signal are calculated, and the asynchronous random signal is obtained. Therefore, the decoder completes the alignment of the effective sampling edge of the sampling clock and the center of the asynchronous random data in a sampling clock period when the data arrives. According to the invention, the accuracy of asynchronous random signal sampling of 1553B bus communication data transmission can be improved.

The invention discloses an antenna feed amplitude-phase quantification method and system based on an improved non-dominated genetic algorithm, and belongs to the technical field of phased-array antennas. The method comprises the following steps: firstly, establishing a far-field pattern calculation model of a phased array, and then optimizing the continuous phase and continuous amplitude of feed in two stages by using an improved non-dominated genetic algorithm: in the first stage, obtaining an optimal phase quantization operation by taking the minimization of a beam pointing error as a target; in the second stage, on the basis of a phase optimization result, the optimal amplitude quantization operation is obtained by taking the maximum sidelobe level of the minimized beam as a target; and finally, a control signal for directly driving the digital phase shifter and the attenuator is generated. According to the method, the amplitude-phase quantization error is systematically controlled by introducing the random quantization interval threshold and error covariance matching technology, the beam pointing precision and the side lobe suppression performance are effectively coordinated by combining elite reservation and adaptive genetic manipulation, and the comprehensive electrical performance of the antenna under the influence of the quantization error is remarkably improved.

Systems, methods, and instrumentalities may be configured for two-level reconfigurable intelligent surface (RIS) channel state information (CSI). A wireless transmit / receive unit may receive a first set of symbols. The WTRU may determine a first RIS CSI based on the first set of symbols, a first RIS CSI estimation matrix, and a first RIS CSI quantization format. The first RIS CSI quantization format may use a uniform distribution of phase quantization points. The WTRU may transmit the first RIS CSI. The WTRU may receive a second set of symbols. The WTRU may determine a second RIS CSI based on the second set of symbols, a second RIS CSI estimation matrix, and a second RIS CSI quantization format. The second RIS CSI quantization format may use a non-uniform distribution of phase quantization points. The WTRU may transmit the second RIS CSI.

Embodiments relate to a method for designing a super lens, the method comprising: selecting a set of super primitives from a library based at least on corresponding phase responses of the set of super primitives; receiving the description of the incident light signal and the target light signal; defining a phase quantization threshold and a maximum number of quantization operations, wherein the phase quantization threshold varies according to the number of quantization operations; generating a phase for each super element of the super lens at least based on the description of the incident light signal and the target light signal by using the IFTA; determining whether a phase generated for each of the super-primitives satisfies a phase quantization threshold, and if the phase quantization threshold is satisfied, performing a quantization operation by quantizing the generated phase to a phase from one of the selected group of super-primitives; and when all quantization operations are completed, phases of all generated super primitives are output.

This invention relates to the field of data transmission technology, specifically to a polarization-selectable variable phase quantization precision reflective array antenna element, antenna, and system. The polarization-selectable variable phase quantization precision reflective array antenna element includes: a reconfigurable resonant structure with four radiating arms; four controllable switches, each disposed on one of the four radiating arms; and a control circuit for independently controlling the on / off state of each controllable switch. By configuring the on / off state combinations of the four controllable switches, switching between co-polarized and cross-polarized reflection modes is possible, and phase modulation with variable phase quantization precision is supported in both polarization modes. This solution, by setting controllable switches on the four radiating arms of the same reconfigurable resonant structure and employing a specific on / off state encoding strategy, achieves a high degree of integration of co-polarized and cross-polarized dual-mode response and 1-bit and 2-bit variable phase modulation functions, overcoming the technical bottleneck that the number of operating modes is positively correlated with the complexity of the element structure.

The application relates to a method and device for tracking an ultra-wideband navigation signal based on a linear frequency modulation spread code, and belongs to the technical field of satellite navigation. The method comprises the following steps: receiving an ultra-wideband navigation signal and multiplying the signal with a local in-phase carrier and a local quadrature carrier respectively to obtain in-phase components and quadrature components; obtaining a local linear frequency modulation spread code which is output after being bound by phase calculation, phase quantization, code generation and signal transmission time delay; obtaining a local pseudo-random spread code which is output after being bound by signal transmission time delay; combining the two types of spread codes to generate three types of spread codes, namely delayed, instant and advanced spread codes, and performing coherent integration on the in-phase components and the quadrature components which are divided into three paths; calculating the phase error and the time delay error of the current received signal and the local carrier according to the coherent integration result; and finally, repeatedly generating the carrier and the two types of spread codes based on error feedback to cyclically realize the tracking of the ultra-wideband navigation signal. The method can reduce the complexity of tracking the ultra-wideband navigation signal.

The invention discloses a system and method for realizing low-jitter and low-spurious oblique waves, and the system comprises a phase processing module, a rate register, a waveform data calculation multiplier, a waveform data bit cutting module, a waveform data symbol conversion adder, a spurious processing module, and a waveform bit width bit cutting module. Wherein the phase processing module comprises a phase accumulator, a phase calculation module, a phase holding register, a descending phase calculation module and a waveform positive and negative phase calculation module, the input end of the phase accumulator is loaded with a frequency control word, and the output end of the phase accumulator is coupled with the phase calculation module; the input end of the phase calculation module is loaded with a phase control word, and the output end of the phase calculation module is respectively coupled with the phase holding register and the descending phase calculation module; through a mathematical model and a phase quantization method, rising edge random jitter and overshoot when a high-speed DAC outputs a square signal are effectively eliminated.

This invention discloses a method for accelerating inference of multimodal large models based on KV Cache low-bit quantization, relating to the fields of artificial intelligence and deep learning. The method includes: a pre-filling stage, in which a subspace is constructed by querying the singular value decomposition of the tensor; iterative quantization with orthogonal constraints applied to the key tensor, and token-by-token quantization of the value tensor, both compressed to 2 bits; treating the weights as complex numbers, and quantizing the phase to the fourth root of unity; inference, layer normalization of the visual encoder is fused; in the decoding stage, token-by-token generation is performed, key-value quantization is carried out in real time and stored in the cache, and attention calculation is completed by combining the history. This invention can significantly reduce memory usage, improve inference throughput, fully utilize the 2-bit storage space, maintain model accuracy, and eliminate the need for additional fine-tuning or calibration data.

Methods, systems, and devices for wireless communications are described. The described techniques may enable a user equipment (UE) to allocate bits of a phase quantization based on a relative gain increase of each of a set of amplitude values. For example, the UE may identify a gain matrix that defines a gain associated with allocating quantities of bits to each amplitude value. The UE may accordingly determine, for each bit of the phase quantization, a respective amplitude value to which the UE may allocate the bit. In some examples, the UE may be configured with multiple gain matrices associated with gain increases for multiple respective noise and source distributions. The UE may therefore select the gain matrix from the multiple configured gain matrices based on one or more noise measurements of the channel.

The invention belongs to the technical field of artificial intelligence, and provides a model quantitative perception training method and device, equipment and a storage medium. The method comprises the following steps: determining initial model parameters of a large language model, mapping the initial model parameters to a complex linear layer, determining a first complex weight of an inverse linear branch and a second complex weight of a complex linear branch in the complex linear layer, and performing phase quantization processing on the first complex weight and the second complex weight to obtain a phase quantization result; and determining a first quantized weight corresponding to the first complex weight and a second quantized weight corresponding to the second complex weight, and updating the first complex weight and the second complex weight according to the first quantized weight and the second quantized weight until a preset convergence condition is met. According to the invention, the precision performance of the model under extremely low bit quantization can be effectively improved.