Unverified50% confidenceFactExact time
低秩分解将m×n权重矩阵近似为A(m×r)与B(r×n)的乘积，存储参数量从m×n降至r×(m+n)
1
Sources
50%
Confidence
Long-term
Relevance
7/19/2026
First Seen
Sources
Related Claims
UnverifiedSVD将任意m×n矩阵A分解为U·Σ·V^T，图像压缩通过只保留前k个最大奇异值实现最优低秩逼近（Eckart-Young定理）69% similarUnverifiedLoRA的数学原理是将权重更新矩阵分解为两个低秩矩阵的乘积，文件体积通常仅几十至几百MB61% similarUnverifiedMLA通过将KV矩阵压缩为低维潜在向量缓存，理论上可将KV Cache显存占用降低至标准MHA的5%至13%61% similarUnverifiedPQ乘积量化通过切分向量、局部聚类、编码替代，可将向量存储体积压缩约8倍（四维示例中从16字节压缩到2字节）60% similarUnverifiedMLA将高维K、V矩阵投影至低秩潜在空间，压缩比可达8-16倍59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/560262API
curl https://kongchang.com/api/v1/knowledge/claims/560262MCP
get_claim(id=560262)