Decoupling Braided Tensor Factors
非负张量分解

非负张量分解一、引言非负张量分解(Nonnegative Tensor Factorization,NTF)是一种基于矩阵分解的多维数据分析方法,它可以将高维数据转化为低维表示,并且能够保留原始数据的主要特征。
NTF 在图像处理、语音识别、信号处理等领域得到了广泛的应用。
二、背景知识1. 张量张量(Tensor)是一种广义的矩阵,它可以表示多维数组。
在机器学习和数据挖掘中,我们经常需要处理高维数据,因此张量成为了非常重要的概念。
2. 非负矩阵分解非负矩阵分解(Nonnegative Matrix Factorization,NMF)是一种常见的降维方法,它可以将一个矩阵分解为两个非负矩阵的乘积。
NMF 能够提取出原始数据中的主要特征,并且具有良好的可解释性。
3. 多维尺度分析多维尺度分析(Multi-Dimensional Scaling,MDS)是一种将高维空间中的点映射到低维空间中的方法。
MDS 可以用于可视化高维数据,并且能够保留原始数据之间的距离关系。
三、非负张量分解的原理1. 目标函数假设我们有一个 $n$ 维张量 $X$,我们希望将其分解为 $r$ 个非负矩阵的乘积,即:$$X \approx \sum_{i=1}^r A_1(:,i) \circ A_2(:,i) \circ \cdots \circA_n(:,i)$$其中 $\circ$ 表示哈达玛积(Hadamard Product),即对应元素相乘。
$A_k(:,i)$ 表示第 $k$ 个矩阵的第 $i$ 列。
我们需要找到一组非负矩阵 $A_1, A_2, \cdots, A_n$,使得它们的乘积能够最好地逼近原始张量 $X$。
为了实现这个目标,我们需要定义一个目标函数:$$\min_{A_k \geq 0} \| X - \sum_{i=1}^r A_1(:,i) \circ A_2(:,i) \circ \cdots \circ A_n(:,i) \|_F^2$$其中 $\|.\|_F$ 表示 Frobenius 范数。
纹理物体缺陷的视觉检测算法研究--优秀毕业论文

摘 要
在竞争激烈的工业自动化生产过程中,机器视觉对产品质量的把关起着举足 轻重的作用,机器视觉在缺陷检测技术方面的应用也逐渐普遍起来。与常规的检 测技术相比,自动化的视觉检测系统更加经济、快捷、高效与 安全。纹理物体在 工业生产中广泛存在,像用于半导体装配和封装底板和发光二极管,现代 化电子 系统中的印制电路板,以及纺织行业中的布匹和织物等都可认为是含有纹理特征 的物体。本论文主要致力于纹理物体的缺陷检测技术研究,为纹理物体的自动化 检测提供高效而可靠的检测算法。 纹理是描述图像内容的重要特征,纹理分析也已经被成功的应用与纹理分割 和纹理分类当中。本研究提出了一种基于纹理分析技术和参考比较方式的缺陷检 测算法。这种算法能容忍物体变形引起的图像配准误差,对纹理的影响也具有鲁 棒性。本算法旨在为检测出的缺陷区域提供丰富而重要的物理意义,如缺陷区域 的大小、形状、亮度对比度及空间分布等。同时,在参考图像可行的情况下,本 算法可用于同质纹理物体和非同质纹理物体的检测,对非纹理物体 的检测也可取 得不错的效果。 在整个检测过程中,我们采用了可调控金字塔的纹理分析和重构技术。与传 统的小波纹理分析技术不同,我们在小波域中加入处理物体变形和纹理影响的容 忍度控制算法,来实现容忍物体变形和对纹理影响鲁棒的目的。最后可调控金字 塔的重构保证了缺陷区域物理意义恢复的准确性。实验阶段,我们检测了一系列 具有实际应用价值的图像。实验结果表明 本文提出的纹理物体缺陷检测算法具有 高效性和易于实现性。 关键字: 缺陷检测;纹理;物体变形;可调控金字塔;重构
Keywords: defect detection, texture, object distortion, steerable pyramid, reconstruction
II
张量分解方程

张量分解方程张量分解方程是一种多维数据分析的统计技术,它用来通过捕获低阶张量中的核心特征,以抽取图像、文本或其他形式的大规模数据。
张量分解可以根据张量中存在的不同特性、不同聚类等来给出定量描述,从而将其应用于知识发现、机器学习、深度学习、计算机视觉等诸多领域。
张量分解方程(Tensor Decomposition)是一类数学模型,通过分解原始张量中存储的高阶特征,从而将庞大的原始张量数据拆解成更加简单的多个张量,这些张量去除部分高阶的特征,而关注低阶的特征,以此来抽取大规模数据的核心特征。
张量分解方程可以将某个原始张量分解为多个低阶张量,分解原理就是对原始张量进行数学变换,使得原始张量中存储的潜在特征可以更加清晰的呈现出来,从而实现从这些低阶张量中提取核心特征的目的。
张量分解方程成功的应用在诸多领域,其中最典型的应用便是知识发现、机器学习和深度学习等。
知识发现利用张量分解可以提取出原始数据集中的潜在特征,从而发现其中的规律;机器学习和深度学习利用张量分解可以在抽取出特征的基础上,训练模型,从而实现计算机视觉等诸多领域的深入研究。
张量分解方程具备多种类型,以不同的变换形式来获取已知的原始张量,大致主要分为非负张量分解(NoT)、非负矩阵分解(NMF)、独立分量分析(ICA)、逐步张量分解(ST)、模型数学分解(MFA)、应用到半监督张量分解(HS-TD)等等。
为了避免因张量分解而产生过拟合,将会引入正则项,实现更加稳定、鲁棒的张量分解,从而提高分析的准确性。
除了引入正则项外,控制张量分解的参数也是减少张量分解过拟合的有效策略,张量分解的参数主要有正则参数、衰减参数、优化次数等,需要结合实际需求加以调节,以保证张量分解的有效性。
尽管利用张量分解可以有效抽取大规模数据中的核心特征,但是由于张量分解涉及到多维数据,相应的计算量也比较大,会耗费较长的时间。
为此,在使用张量分解时,采取分布式计算的策略,可以减少计算量,有效提升计算效率。
扩张因果卷积模型

扩张因果卷积模型
扩张因果卷积模型(Dilated Causal Convolutional,DCC)是一种特殊的卷积神经网络结构,由Oord A V D在2016年提出。
这种模型的设计目的是为了解决因果卷积中的问题,即通过增加感受野来扩大卷积核的应用范围。
在因果卷积中,卷积核只能读取当前信息和历史信息,这限制了其对时间序列数据的处理能力。
为了解决这个问题,扩张因果卷积模型通过有规律的跳过部分输入来使卷积核可以应用于大于卷积核本身长度的区域,从而增加了感受野。
具体来说,扩张卷积通过在原始卷积核周围增加一些零值,从而生成更大的卷积核。
这种扩大卷积核的方式可以使模型在层数不大的情况下有非常大的感受野。
在示意图中,可以看到卷积感受野扩大了1、2、4、8倍。
使用扩张因果卷积模型,可以使模型学习数据不同时间间隔的特征,提取数据的空间特征。
由于减少了由于下采样操作而导致的分辨率或覆盖率损失,模型可以更有效地处理数据。
ai常用的底层算子

ai常用的底层算子AI常用的底层算子一、卷积算子(Convolution)卷积算子是深度学习中常用的底层算子之一。
它通过对输入数据与卷积核进行卷积操作,实现特征的提取和图像的处理。
卷积算子可以应用于图像处理、语音识别、自然语言处理等领域。
二、池化算子(Pooling)池化算子用于减小特征图的尺寸和参数的数量,从而降低计算复杂度并增强模型的鲁棒性。
常见的池化算子有最大池化(Max Pooling)和平均池化(Average Pooling),它们分别通过选取最大值或平均值来减小特征图的尺寸。
三、激活函数(Activation Function)激活函数是神经网络中常用的底层算子之一。
它引入非线性因素,使神经网络能够学习复杂的非线性关系。
常见的激活函数有Sigmoid函数、ReLU函数、Tanh函数等。
四、批归一化算子(Batch Normalization)批归一化算子用于提高神经网络的训练速度和稳定性。
它通过对每一层的输入进行归一化处理,从而加速收敛过程并减少梯度消失或梯度爆炸的问题。
五、全连接算子(Fully Connected)全连接算子是神经网络中常用的底层算子之一。
它将输入数据与权重矩阵进行矩阵相乘,并加上偏置项,得到最终的输出。
全连接算子可以实现特征的组合和高维向量的映射。
六、损失函数(Loss Function)损失函数是神经网络中用于度量模型预测与真实值之间差异的函数。
常见的损失函数有均方误差(MSE)、交叉熵损失函数等。
损失函数的选择会影响模型的训练效果和收敛速度。
七、优化算法(Optimization Algorithm)优化算法用于调整神经网络中的参数,使得损失函数达到最小值。
常见的优化算法有随机梯度下降(SGD)、Adam、Adagrad等。
优化算法的选择会影响模型的训练速度和性能。
八、循环神经网络(Recurrent Neural Network, RNN)循环神经网络是一种具有循环连接的神经网络结构,用于处理序列数据。
张量链式分解

张量链式分解引言张量链式分解(Tensor Chain Decomposition)是一种用于对多维数据进行降维和特征提取的技术。
它广泛应用于各个领域,包括图像处理、语音识别、推荐系统等。
本文将介绍张量链式分解的基本概念、原理、方法和应用,并探讨其优缺点以及未来的发展方向。
张量基础知识在深入讲解张量链式分解之前,我们先来了解一些张量的基础知识。
1. 张量的定义首先,我们需要明确张量(Tensor)的概念。
在数学和物理学中,张量是一种多维数组或矩阵的推广。
例如,0阶张量是标量(Scalar),1阶张量是向量(Vector),2阶张量是矩阵(Matrix),以此类推。
2. 张量的表示张量可以用多种方式进行表示,包括矩阵、数组和张量的分块表示等。
其中,张量的分块表示可以简化数据的处理和计算。
3. 张量的运算和矩阵类似,张量也支持多种运算,包括加法、减法、乘法、求逆等。
这些运算可以用于数据的变换和处理。
张量链式分解的基本概念接下来,我们将介绍张量链式分解的基本概念。
1. 张量链的定义张量链(Tensor Chain)是指由多个张量组成的序列。
每个张量可以有不同的维度和大小,但它们的维度必须满足一定的连续性条件。
2. 张量链的分解张量链的分解是将复杂的张量链表示为一系列低阶张量的乘积的过程。
分解后的低阶张量包含了原始张量链的信息,并可以用于降维和特征提取。
3. 张量链式分解的原理张量链式分解的原理是基于张量的秩的概念。
张量的秩是指表示张量的低阶张量的个数。
通过适当的张量链分解,可以降低张量的秩,从而实现降维和数据压缩。
张量链式分解的方法张量链式分解有多种方法和算法,下面介绍其中几种常见的方法。
1. CP分解(CANDECOMP/PARAFAC)CP分解是一种基于多线性代数的张量分解方法。
它将一个高阶张量表示为多个低阶张量的和。
CP分解具有数学上的优良性质和可解释性,被广泛应用于数据分析和模型简化。
2. Tucker分解Tucker分解是一种将一个高阶张量分解为一系列核张量与模态矩阵的乘积的方法。
separable convolution公式

separable convolution公式Separable Convolution Formula: An IntroductionConvolutional neural networks (CNNs) have revolutionized the field of computer vision, enabling remarkable progress in various image processing tasks. Among the essential components of CNNs, the separable convolution operation has gained significant attention due to its efficiency and effectiveness in reducing computational complexity while preserving accuracy.The separable convolution is a technique that decomposes a traditional 2D convolution into two consecutive operations: a spatial convolution and a depthwise convolution. This decomposition can be denoted by the following formula:output[x, y, z] = ∑∑∑ kernel_spatial[i, j] * kernel_depthwise[c, z] * input[x+i, y+j, c]In the above formula, "output" represents the resulting feature map, "kernel_spatial" denotes the spatial convolution kernel, "kernel_depthwise" represents the depthwise convolution kernel, and "input" refers to the input feature map. The variables (x, y, z) iterate over the spatial dimensions (width, height, depth) of the output feature map, while (i, j, c) iterate over the respective dimensions of the convolution kernels and input feature map.The separable convolution takes advantage of the fact that most CNNs learn spatial and channel-wise filters independently. By separating the 2D convolution into two distinct operations, the separable convolution significantly reduces the number of parameters and computations required.The spatial convolution involves convolving each channel of the input feature map with its respective spatial kernel. This operation helps capture spatial patterns and local correlations within each channel independently. On the other hand, the depthwise convolution applies a 1x1 point-wise convolution across each channel of the spatially convolved feature map. This allows cross-channel interactions and provides a way to combine the spatial filters.Compared to the standard 2D convolution, separable convolution offers several advantages. First, it greatly reduces computational complexity, resulting in faster inference and training times. Second, it alleviates overfitting by reducing the capacity of the model, preventing excessive memorization. Lastly, the separable convolution can achieve similar or even better performance compared to the standard convolution in many scenarios.In conclusion, the separable convolution formula provides a powerful technique that enhances the efficiency and effectiveness of convolutional neural networks. By decomposing the traditional 2D convolution into spatial and depthwise convolutions, separable convolution reduces computational complexity while maintaining accuracy. This technique has become a fundamental building block in modern CNN architectures, enabling remarkable advancements in image processing tasks.。
tensor相似度损失计算公式

文章标题:深度学习中的tensor相似度损失计算公式探究在深度学习中,tensor相似度损失计算公式是一个非常重要的概念。
本文将从浅入深地探讨这一主题,帮助您更好地理解和运用它。
1. 介绍让我们来了解一下什么是tensor相似度损失。
在深度学习中,我们经常需要比较两个张量(tensor)的相似度,以评估它们之间的关系。
而损失函数则是深度学习模型中非常重要的一部分,它用于衡量模型输出与真实数值之间的差异。
tensor相似度损失计算公式可以帮助我们计算模型输出张量与真实标签张量之间的相似度损失,进而指导模型优化。
2. 相似度计算方法针对不同类型的张量,我们可以使用不同的相似度计算方法。
常用的包括欧氏距离、余弦相似度、交叉熵等。
在深度学习中,我们经常会用到欧氏距离和余弦相似度作为相似度计算的方法。
这两种方法分别适用于不同类型的数据,欧氏距离更适用于数值型数据,而余弦相似度更适用于向量型数据。
在实际应用中,我们需要根据数据的特点来选择合适的相似度计算方法。
3. tensor相似度损失计算公式针对tensor的相似度损失计算公式,我们通常会使用欧氏距离或余弦相似度的计算方法。
对于欧氏距离,公式可以表示为:\[ \sqrt{\sum_{i=1}^{n}(A_i - B_i)^2} \]其中,\( A_i \) 和 \( B_i \) 分别表示两个张量中对应位置的元素。
这个公式能够计算出两个张量之间的距离,距离越小表示相似度越高。
4. 深入理解在实际应用中,tensor相似度损失计算公式是深度学习模型优化中不可或缺的一环。
通过合理选择相似度计算方法和损失计算公式,我们能够更好地指导模型学习,提高模型的准确性和泛化能力。
而对于不同的任务和数据类型,我们需要灵活选择相似度计算方法和调整损失计算公式,以达到最佳的效果。
5. 个人观点在我的理解中,tensor相似度损失计算公式是深度学习中的一项重要技术,它能够帮助模型理解输入数据的特征,优化模型的学习过程。
- 1、下载文档前请自行甄别文档内容的完整性,平台不提供额外的编辑、内容补充、找答案等附加服务。
- 2、"仅部分预览"的文档,不可在线预览部分如存在完整性等问题,可反馈申请退款(可完整预览的文档不适用该条件!)。
- 3、如文档侵犯您的权益,请联系客服反馈,我们会尽快为您处理(人工客服工作时间:9:00-18:30)。
a rX iv:mat h /12199v1[mat h.QA ]2Dec2Decoupling Braided Tensor Factors ∗Gaetano Fiore,1,2Harold Steinacker 3,Julius Wess 3,41Dip.di Matematica e Applicazioni,Fac.di Ingegneria Universit`a di Napoli,V.Claudio 21,80125Napoli 2I.N.F.N.,Sezione di Napoli,Complesso MSA,V.Cintia,80126Napoli 3Sektion Physik,Ludwig-Maximilian Universit¨a t,Theresienstraße 37,D-80333M¨u nchen 4Max-Planck-Institut f¨u r Physik F¨o hringer Ring 6,D-80805M¨u nchen Abstract We briefly report on our result [9]that the braided tensor product algebra of two module algebras A 1,A 2of a quasitriangular Hopf al-gebra H is equal to the ordinary tensor product algebra of A 1with a subalgebra isomorphic to A 2and commuting with A 1,provided thereexists a realization of H within A 1.As applications of the theorem we consider the braided tensor product algebras of two or more quantum group covariant quantum spaces or deformed Heisenberg algebras.1Introduction and main theoremAs is well known,given two associative unital algebras A1,A2(over thefield C,say),there is an obvious way to build a new algebra A which is as a vector space the tensor product A=A1⊗C A2of the two vector spaces(over the samefield)and has a product law such that A1⊗1and1⊗A2are subalgebrasisomorphic to A1and A2respectively:one just completes the product law by postulating the trivial commutation relations(1⊗a2)(a1⊗1)=(a1⊗1)(1⊗a2)(1) for any a1∈A1,a2∈A2.The resulting algebra is the ordinary tensor product algebra.With a standard abuse of notation we shall denote in the sequel a1⊗a2by a1a2for any a1∈A1,a2∈A2;consequently(1)becomesa2a1=a1a2.(2) If A1,A2are module algebras of a Lie algebra g,and we require A to be too,then(2)has no alternative,because any g∈g acts as a derivation on the (algebra as well as tensor)product of any two elements,or,in Hopf algebra language,because the coproduct∆(g)=g(1)⊗g(2)(at the rhs we have used Sweedler notation)of the Hopf algebra H≡U g is cocommutative.In this paper we shall work with right-module algebras(instead of left ones),and denote by⊳:(a i,g)∈A i×H→a i⊳g∈A i the right action;the reason is that they are equivalent to left comodule algebras,which are used in much of the literature.In Ref.[9]we give also the corresponding formulae for the left module algebras.We recall that a right action⊳:(a,g)∈A×H→a⊳g∈A by definition fulfillsa⊳(gg′)=(a⊳g)⊳g′,(3)(aa′)⊳g=(a⊳g(1))(a′⊳g(2)).(4) If we take as Hopf algebra H a quasitriangular noncocommutative one like the quantum group U q g,as A i some H-module algebras,and we require A to be a H-module algebra too,then(2)has to be replaced by one of the formulaea2a1=(a1⊳R(1))(a2⊳R(2)),(5)a2a1=(a1⊳R−1(2))(a2⊳R−1(1)).(6)2This yields instead of A two different braided tensor product algebras[10,11], which we shall call A+=A1⊗−A2respectively.Here R≡R(1)⊗R(2)∈H+⊗H−denotes the so-called universal R-matrix of H≡[6],R−1its inverse,and H±denote the Hopf positive and negative Borel subalgebras of H.If in particular H is triangular,then R−1=R21, A+=A−,and one has just one braided tensor product algebra.In any case, both A+and A−go to the ordinary tensor product algebra A in the limit q→1,because in this limit R→1⊗1.The braided tensor product is a particular example of a more general no-tion,that of a crossed(or twisted)tensor product[1]of two unital associative algebras.In view of(5)or(6)studying representations of A±is a more difficult task than just studying the representations of A1,A2and taking their tensor products.The degrees of freedom of A1,A2are so to say“coupled”.One might ask whether one can“decouple”them by a transformation of genera-tors.As shown in Ref.[9],the answer is positive if there respectively exists an algebra homomorphismϕ+1or an algebra homomorphismϕ−1ϕ±1:A1>⊳H±→A1(7) acting as the identity on A1,namely for any a1∈A1ϕ±1(a1)=a1.(8) (Here A1>⊳H±denotes the cross product between A1and H±).In other words,this amounts to assuming thatϕ+1(H+)[resp.ϕ−1(H−)]provides a realization of H+(resp.H−)within A1.In this report we summarize the main results of Ref.[9].The basic one isTheorem1[9].Let{H,R}be a quasitriangular Hopf algebra and H+,H−be Hopf subalgebras of H such that R∈H+⊗H−.Let A1,A2be respectively a H+-and a H−-module algebra,so that we can define A+as in(5),and ϕ+1be a homomorphism of the type(7),(8),so that we can define the map χ+:A2→A+byχ+(a2):=ϕ+1(R(1))(a2⊳R(2)).(9) Alternatively,let A1,A2be respectively a H−-and a H+-module algebra,so that we can define A−as in(6),andϕ−1be a homomorphism of the type(7), (8),so that we can define the mapχ−:A2→A−byχ−(a2):=ϕ−1(R−1(2))(a2⊳R−1(1)).(10)3In either caseχ±are then injective algebra homomorphisms and[χ±(a2),A1]=0,(11) namely the subalgebras˜A±2:=χ±(A2)≈A2commute with A1.Moreover A±=A1⊗˜A±2.The last equality means that A±are respectively equal to the ordinary tensor product algebra of A1with the subalgebras˜A±2⊂A±,which are isomorphic to A2!χ+,χ−will be called”unbraiding”maps.We recall the content of the hypotheses stated in the theorem.The alge-bra A1>⊳H±as a vector space is the tensor product A1⊗C H±,as an algebra it has subalgebras A1⊗1,1⊗H and has cross commutation relationsa1g=g(1)(a1⊳g(2)),(12) for any a1∈A1and g∈H±.ϕ±1being an algebra homomorphism means that for anyξ,ξ′∈A1>⊳H±ϕ±1(ξξ′)=ϕ±1(ξ)ϕ±1(ξ′).Applyingϕ±1to both sides of(12)wefind aϕ±(g)=ϕ±(g(1))(a⊳g(2)).Of course,we can use the above theorem iteratively to completely unbraid the braided tensor product algebra of an arbitrary number M of copies of A1.We end up withCorollary1If A1is a(right-)module algebra of the Hopf algebra H and there exists an algebra homomorphismϕ+1of the type(7),(8),then there is an algebra isomorphism.A1⊗+A1 M times ≈A1⊗...⊗A1M times.(13)An analogous claim holds for the second braided tensor product if there exists a mapϕ−1.2The unbraiding under the∗-structuresA+(as well as A−)is a∗-algebra if H is a Hopf∗-algebra,A1,A2are H-module∗-algebras(we shall use the same symbol∗for the∗-structure on all algebras H,A1,etc.),andR∗=R−1(14)4(here R∗means R(1)∗⊗R(2)∗).In the quantum group case(14)requires |q|=1.Under the same assumptions also A1>⊳H is a∗-algebra.Ifϕ±1exist settingϕ′1±:=∗◦ϕ±1◦∗we realize that alsoϕ′1±are algebra homomorphisms of the type(7),(8).If such homomorphisms are uniquely determined,we conclude thatϕ1±are∗-homomorphisms.More generally,one may be able to chooseϕ1±as∗-homomorphisms.How do the correspondingχ±behave under∗?Proposition1[9].Assume that the conditions of Theorem1for defin-ingχ+(resp.χ−)are fulfilled.If R∗=R−1andϕ+1(resp.ϕ−1)is a ∗-homomorphism thenχ+(resp.χ−)is,too.Consequently,A1,˜A±2are closed under∗.3ApplicationsIn this section we illustrate the application of Theorem1and Corollary1 to some algebras H,A i for which homomorphismsϕ±1are known.H will be the quantum group U q sl(N)or U q so(N),and A1is the U q sl(N)-or U q so(N)-covariant Heisenberg algebra(Section3.1.),the U q so(N)-covariant quantum space/sphere(Section3.2.).In Ref.[9]we have treated also the U q so(3)-covariant q-fuzzy sphere.As generators of H it will be convenient in either case to use the Faddeev-Reshetikhin-Takhtadjan(FRT)generators[7]L+a l∈H+and L−a l∈H−.They are related to R byL+a l:=R(1)ρa l(R(2))L−a l:=ρa l(R−1(1))R−1(2),(15) whereρa l(g)denote the matrix elements of g∈U q g in the fundamental N-dimensional representationρof U q g.In fact they provide,together with the square roots of the elements L±i i,a(overcomplete)set of generators of U q g.3.1.Unbraiding‘chains’of braided Heisenberg alge-brasIn this subsection we consider the braided tensor product of M≥2copies of the U q g-covariant deformed Heisenberg algebras Dǫ,g,g=sl(N),so(N). Such algebras have been introduced in Ref.[13,14,2].They are unital associative algebras generated by x i,∂j fulfilling the relationsP a ij hk x h x k=0,P a ij hk∂j∂i=0,∂i x j=δi j+(qγˆR)ǫjk ih x h∂k,(16)5whereγ=q12 denotes the rank of so(N).We shall enumerate the different copies of Dǫ,g by attachingto them an additional Greek index,e.g.α=1,2,...,M.The prescription(6)gives the following“cross”commutation relations between their respective generators(α<β).xα,i xβ,j=ˆR ijhkxβ,h xα,k,∂α,i∂β,j=ˆR kh ji∂β,h∂α,k,∂α,i xβ,j=ˆR−1jhik xβ,k∂α,h,∂β,i xα,j=ˆR jhikxα,k∂β,h.(18)Algebra homomorphismsϕ1:A1>⊳H→A1,for H=U q g and A1equal to(a suitable completion of)Dǫ,g have been constructed in Ref.[8,5].This is the q-analog of the well-known fact that the elements of g can be realized as“vectorfields”(first order differential operators)on the corresponding g-covariant(undeformed)space,e.g.ϕ1(E i j)=x i∂j−1ϕ1(L−i j)in terms of x1,i,∂1,a for U q sl(2),U q so(3)has been given in Ref.[9]. For different values of N it can be found from the results of Ref.[8,5]by passing from the generators adopted there to the FRT generators.By completely analogous arguments one determines the alternative un-braiding procedure for the braided tensor product stemming from prescrip-tion(5).A1>⊳H is a∗-algebra and the mapϕ1is a∗-homomorphism both for q real and|q|=1.Butϕ±1are∗-homomorphisms only for|q|=1.In the latter case the∗-structure of A1is(x i)∗=x i,(∂i)∗=−q±N g kh g ki∂h(22) Applying Proposition1in the latter case wefind that∗maps A1as well as each of the commuting subalgebras˜A±i into itself.3.2.Unbraiding‘chains’of braided quantum Euclideanspaces or spheresIn this section we consider the braided tensor product of M≥2copies of the quantum Euclidean space R N q[7](the U q so(N)-covariant quantum space),i.e.of the unital associative algebra generated by x i fulfilling the relations(16)1,or of the quotient space of R N q obtained by setting r2:=x i x i=1[the quantum(N−1)-dimensional sphere S N−1q ].(Thus,these willbe subalgebras of the Heisenberg algebras D+,so(N),D+,so(N)considered in the previous subsection).Again,the multiplet(x i)carries the fundamental N-dimensional representationρof U q so(N).As before,we shall enumerate the different copies of the quantum Euclidean space or sphere by attaching an additional Greek index to them,e.g.α=1,2,...,M.The prescription(6) gives the cross commutation relations(18)1.According to Ref.[3],to defineϕ±1(for q=1)one actually needs a slightlyenlarged version of R N q(or S N−1q ).One has to introduce some new generators√2,together with their inverses(√2 ,r2n coincides with r2,whereas for odd N r20=(x0)2,so we are adding also(x0)−1as a new generator).In fact,the com-mutation relations involving these new generators can befixed consistently,7and turn out to be simply q-commutation relations.r plays the role of‘de-formed Euclidean distance’of the generic‘point’(x i)of R N q from the‘origin’; r a is the‘projection’of r on the‘subspace’x i=0,|i|>a.In the previous equation g hk denotes the‘metric matrix’of SO q(N),g ij=g ij=q−ρiδi,−j, which is a SO q(N)-isotropic tensor and a deformation of the ordinary Eu-clidean metric.Here,(ρi):=(n−12,0,−12−n)for N odd,(ρi):=(n−1,...,0,0,...,1−n)for N even.g ij is related to the traceprojector appearing in(28)by P t ijkl =(g sm g sm)−1g ij g kl.The extension ofthe action of H to these extra generators is uniquely determined by the con-straints the latter fulfil.In the case of even N one needs to include also the FRT generators L+11,L−11(which are generators of H)among the generators of A1.In appendix3.2.we recall the explicit form ofϕ±1in the present case. Note that the mapsϕ±1have no analog in the“undeformed”case(q=1), because A1is abelian,whereas H is not.The unbraiding procedure is recursive.Thefirst step consists of using the homomorphismϕ±1found in Ref.[3]to unbraid thefirst copy from the others.Following Theorem1,we perform the change of generators(19)1, (20)in A−.In view of formula(29)we thusfindy1,i:=x1,i,yα,i:=g ih[µ1h,x1,k]q g kj xα,j,α>1.(24) The suffix1inµ1a means that the special elementsµa defined in(30)must be taken as elements of thefirst copy.In view of(30)we see that g ih[µ1h,x1,k]q g kj are rather simple polynomials in x i and r−1a,homogeneous of total degree1 in the coordinates x i and r ing the results given in the appendix we give now the explicit expression of(24)2for N=3:yα,−=−qhγ1rq(q+1)1q(q+1)hγ1rx+xα,0−1q−1/√The alternative unbraiding procedure for the braided tensor product al-gebra stemming from prescription(5)arises by iterating the change of gen-eratorsy′M,i:=x M,i y′α,i:=ϕM(L+i j)xα,j=g ih[¯µM h,x M,k]q−1g kj xα,j,(26)α<M.The special elements¯µa are defined in(30),and suffix M meansthat we must take¯µa as an element of the M-th copy of R N q(or S N−1q ).y M,i≡x M,i commutes with y1,i,...,y M−1,i.When|q|=1,by a suitable choice(32)ofγ1,¯γ1,as well as of the other free parametersγa,¯γa appearing in the definitions ofϕ±for N>3,one can makeϕ±into∗-homomorphisms.Applying Proposition1in the latter case wefind that∗maps A1as well as each of the commuting subalgebras˜A±i into itself.The braid matrixˆR is related to R byˆR ijhk≡R ji hk:=(ρj h⊗ρi k)R.With the indices’convention described in sections3.2.,3.1.ˆR is given byˆR=q−1AB −xBA .The images of ϕ−(resp.ϕ+)on the negative (resp.positive)FRT generators readϕ−(L −i j )=g ih [µh ,x k ]q g kj ,ϕ+(L +i j )=g ih [¯µh ,x k ]q −1g kj ,(29)whereµ0=γ0(x 0)−1¯µ0=¯γ0(x 0)−1for N odd,µ±1=γ±1(x ±1)−1L ∓11¯µ±1=¯γ±1(x ±1)−1L ±11for N even,µa =γa r −1|a |r −1|a |−1x −a ¯µa =¯γa r −1|a |r −1|a |−1x −a otherwise,(30)and γa ,¯γa ∈C are normalization constants fulfilling the conditions γ0=−q −12h −1for N odd,γ1γ−1= −q −1h −2k −2¯γ1¯γ−1= −qh −2k −2for N odd,for N even,γa γ−a =−q −1k −2ωa ωa −1¯γa ¯γ−a =−qk −2ωa ωa −1for a >1.(31)Here k :=q −1/q ,ωa :=(q ρa +q −ρa ).Incidentally,for odd N one can choose the free parameters γa ,¯γa in such a way that ϕ+,ϕ−can be ‘glued’into an algebra homomorphism ϕ:R N q >⊳U q so(N )→R N q [3].When |q |=1,the ∗-structure is given by (x i )∗=x i [see (22)].It turns out that ϕ±are ∗-homomorphisms if,in addition,γ∗±1=−γ±1if N even γ∗a =−γa 1if a <0q −2if a >0otherwise.(32)References[1]A.Van Daele and S.Van Keer,Compositio Mathematica 91,201(1994).A.Borowiec,W.Marcinek,“On crossed product of algebras”,math-ph/0007031,and references therein;“Hopf modules and their duals”,math.QA/0007151.[2]U.Carow-Watamura,M.Schlieker,S.Watamura,Z.Physik C 49(1991)439.10[3]B.L.Cerchiai,G.Fiore,J.Madore,“Geometrical Tools forQuantum Euclidean Spaces”,to appear in Commun.Math.Phys., math.QA/0002007[4]B.L.Cerchiai,J.Madore,S.Schraml,J.Wess,Eur.Phys.J.C16(2000),169-180.[5]C.-S.Chu,B.Zumino,“Realization of vectorfields fro quantum groupsas pseudodifferential operators on quantum spaces”,Proc.XX Int.Conf.on Group Theory Methods in Physics,Toyonaka(Japan),1995,and q-alg/9502005.[6]V.Drinfeld,“Quantum groups,”in I.C.M.Proceedings,Berkeley,(1986)p.798.[7]L.D.Faddeev,N.Y.Reshetikhin,L.Takhtadjan,Leningrad Math.J.1(1990),193.[8]G.Fiore,Commun.Math.Phys.169(1995),475-500.[9]G.Fiore,H.Steinacker,J.Wess“Unbraiding the braided tensor prod-uct”,Preprint00-30Dip.Matematica e Applicazioni,Universit`a di Napoli,math/0007174.[10]A.Joyal,R.Streat,Braided Monoidal Categories,Mathematics Reports86008,Macquarie University,1986.[11]S.Majid,Int.J.Mod.Phys.A5,1(1990);J.Algebra130,17(1990);Lett.Math.Phys.22,167(1991);J.Algebra163,191(1994).For a review:S.Majid,Foundations of Quantum Groups,Cambridge Univ.Press(1995);and references therein.[12]O.Ogievetsky,Lett.Math.Phys.24(1992),245.[13]W.Pusz,S.L.Woronowicz,Rep.Math.Phys.27(1989),231.[14]J.Wess,B.Zumino,Nucl.Phys.(Proc.Suppl.)18B(1990)302.11。