• <tr id="yyy80"></tr>
  • <sup id="yyy80"></sup>
  • <tfoot id="yyy80"><noscript id="yyy80"></noscript></tfoot>
  • 99热精品在线国产_美女午夜性视频免费_国产精品国产高清国产av_av欧美777_自拍偷自拍亚洲精品老妇_亚洲熟女精品中文字幕_www日本黄色视频网_国产精品野战在线观看 ?

    An ?-Step Modified Augmented Lagrange Multiplier Algorithm for Completing a Toeplitz Matrix

    2019-10-16 01:44:40WENRuiping溫瑞萍LIShuzhen李姝貞
    應(yīng)用數(shù)學(xué) 2019年4期

    WEN Ruiping(溫瑞萍),LI Shuzhen(李姝貞)

    ( Key Laboratory of Engineering & Computing Science,Shanxi Provincial Department of Education /Department of Mathematics,Taiyuan Normal University,Jinzhong 030619,China)

    Abstract: Based on the modified augmented Lagrange multiplier (MALM) algorithm for Toeplitz matrix completion (TMC) proposed by WANG et al.(2016),we put forward an accelerated technique to MALM algorithm,which will reduce the extra load coming from data communication.It is drawn that an ?-step modified augmented Lagrange multiplier algorithm.Meanwhile,we demonstrate the convergence theory of the new algorithm.Finally,numerical experiments show that the ?-step modified augmented Lagrange multiplier(?-MALM) algorithm is more effective than the MALM algorithm.

    Key words: Toeplitz matrix;Matrix completion;Augmented Lagrange multiplier;Data communication

    1.Introduction

    Completing an unknown low-rank or approximately low-rank matrix from a sampling of its entries is a challenging problem applied in many fields of information science.For example,machine learning[1?2],control theory[11],image inpainting[4],computer vision[14]etc.Matrix completion (MC) problem first introduced by Cand`es and Recht[6]is from portion of the observed matrix elements to fill a low-rank matrix as precisely as possible,it is a hot spot of researching in recent years.The famous recommendation system of the Netflix[3]is the typical application of matrix completion.Mathematical language is expressed as follows:

    where the matrixM∈Rm×nis the underlying matrix to be reconstructed,? ?{1,2,···,m}×{1,2,···,n} represents the random subset of indices for the known entries to the sampling matrixM,andP?is associated sampling orthogonal projection operator on?.The purpose of the optimization problem is to reduce the rank of matrix obtained by filling missing elements as far as possible.In other words,the model (1.1) optimizes the structure of the matrix.

    Although matrix completion needs solving the global solution of the non-convex objective problem,there are still many effective algorithms which are applicable to some specific matrix.Many researchers have suggested that the most of a low-rank matrix can be accurately completed according to the known entries at specified accuracy.In a low dimensional linear subspace,the meaning of precision is that we can realize the reasonable matrix completion under the hypothesis of the lowest-rank data matrix.

    However,the optimization problem (1.1) is an NP-hard problem,in theory and practice,the computational complexity of the existing algorithms is a double exponential function about the matrix dimension.As a result,the computations are cumbersome and the data take up a large amount of computer memory.Therefore,the solving of the problem (1.1) is usually expensive by the existing algorithms.Hence,using the relation between the rank and the nuclear norm of a matrix,Cand`es and Recht[6]made the equivalent form of the problem(1.1) as follows:

    whereM∈Rm×nis the underlying matrix,denotes thek-th largest singular value of ther-rank matrixA∈Rm×n,namelyσ1≥σ2≥···≥σk≥···≥σr >0.The problem (1.1) is transformed into a convex optimization problem which is easy to solve.

    Academic circles have made abundant research results in solving the optimization problem (1.2) such as the augmented Lagrange multiplier (ALM) algorithm[9],the singular value theresholding(SVT)algorithm as well as its variants[5,8,19],the accelerated proximal gradient(APG)algorithm[15],and the modified augmented Lagrange multiplier(MALM)algorithm[17]etc.On the practical problems,the sampling matrix often has a special structure,for instance,the Toeplitz and Hankel matrices.Many scholars have conducted in-depth research on the special structure,property and application of the Toeplitz and Hankel matrix in recent years[10,12?13,16,18].As we well know,ann×nToeplitz matrix can be expressed as the following form:

    which is determined by 2n?1 entries.

    As for MALM algorithm,the intuitive idea is to switch iteration matrix into Toeplitz structure at each step by an operator.To implement this idea,a heavy data has to be moved at each iterate step.However,there is a cost,sometimes relatively great,associated with the moving of data.The control of memory traffic is crucial to performance in many computers.

    In order to reduce the traffic jam of data,an?-step modified ALM algorithm is proposed in this paper.Compared with the MALM algorithm,the new algorithm saves computation cost and reduces data communication.Two aspects are taken into account,which result in a more practical or economic implementing.The new algorithm not only overcomes the slowness produced by the computing of singular value decomposition in ALM algorithm,but also saves the data congestion caused by the data communication in MALM algorithm.Compared with the CPU of the MALM algorithm,we can see that the CPU of the?-MALM algorithm is reduced to 45%.

    Here are some of the necessary notations and preliminaries.Rm×ndenotes the set ofm×nreal matrix.The nuclear norm of a matrixAis denoted by‖A‖?,the Frobenius norm by‖A‖F(xiàn)is the sum of absolute values of the matrix entries ofA.ATis used to denote the transpose of a matrixA∈Rn×n,rank(A) stands for the rank ofAand tr(A) represents the trace ofA.The standard inner product between two matrices is denoted by〈X,Y〉=tr(XTY).? ?{?n+1,···,n?1}is the indices set of observed diagonals of a Toeplitz matrixM∈Rn×n,ˉ?is the complementary set of?.For a Toeplitz matrixA∈Rn×n,the vector vec(A,l)denotes a vector reshaped from thel-th diagonal ofA,l=?n+1,···,n?1,P?is the orthogonal projector on?,satisfying

    Definition 1.1(Singular value decomposition(SVD)) The singular value decomposition of a matrixA∈Rm×nofr-rank is defined as follows:

    whereU∈Rm×randV∈Rn×rare column orthonormal matrices,σ1≥σ2≥···≥σr >0.

    Definition 1.2(Singular value thresholding operator)[5]For eachτ≥0,the singular value thresholding operatorDτis defined as follows:

    whereA=UΣrVT∈Rm×n,{σi?τ}+=

    Definition 1.3The matrices

    are called the basis of the Toeplitz matrix space.

    It is clear that a Toeplitz matrixT∈Rn×n,shown in (1.3),can be rewritten as a linear combination of these basis matrices,that is,

    Definition 1.4(Toeplitz structure smoothing operator) For any matrixA=(aij)∈Rn×n,the Toeplitz structure mean operatorTis defined as follows:

    The rest of the paper is organized as follows.After we review briefly the ALM,MALM algorithms and the dual approach,an?-step modified ALM algorithm will be proposed in Section 2.Next,the convergence analysis is given in Section 3.Then the numerical results are provided to show the effectiveness of the?-MALM in Section 4.Finally,the paper ends with a conclusion in Section 5.

    2.Relative Algorithms

    Since the matrix completion problem is closely connected to the robust principal component analysis (RPCA) problem,then it can be formulated in the same way as RPCA,an equivalent problem of (1.2) can be considered as follows.In terms of estimating the lowdimensional subspace,the purpose of the mathematical model is to find a low-rank matrixA∈Rm×n(as long as the error matrixEis sufficiently sparse,relative to the rank ofA)to minimize the difference between matrixAandM,generating the following constraint optimization problem model:

    whereEwill compensate for the unknown entries ofM,the unknown entries ofM∈Rm×nare simply set as zeros.AndP?:Rm×n→Rm×nis a linear operator that keeps the entries in?unchanged and sets those outside?(say,in) zeros.Then we introduce the algorithm for solving problem (2.1).

    ⅠThe augmented Lagrange multiplier (ALM) algorithm

    It is famous that partial augmented Lagrangian function of the problem (2.1) is Hence,the augmented Lagrange multiplier (ALM) algorithm[9]is designed as follows.

    Algorithm 2.1Step 0 Give?,sampled matrixD=P?(M),μ0>0,ρ >1.Give also two initial matricesY0=0,E0=0.k:=0;

    Step 1 Compute the SVD of the matrix (D?Ek+μk?1Yk),

    Step 2 Set

    SolveEk+1=arg

    Step 3 If‖D?Ak+1?Ek+1‖F(xiàn)/‖D‖F(xiàn)

    Step 4 SetYk+1=Yk+μk(D?Ak+1?Ek+1).Ifμk‖Ek+1?Ek‖F(xiàn)/‖D‖F(xiàn)

    RemarkIt is reported that the ALM algorithm performs better both in theory and algorithms than the others that with a Q-linear convergence speed globally.It is of much better numerical behavior,and it is also of much higher accuracy.However,the algorithm has a disadvantage of the penalty function:the matrix sequences{Xk} generated by Algorithm 2.1 are not feasible.Hence,the accepted solutions are not feasible.

    ⅡThe dual algorithm

    The dual algorithm proposed in [7]tackles the problem (2.1) via its dual.That is,one first solves the dual problem

    for the optimal Lagrange multiplierY,where

    A steepest ascend algorithm constrained on the surface{Y|J(Y)=1}can be adopted to solve(2.3),where the constrained steepest ascend direction is obtained by projectingMonto the tangent cone of the convex body{Y|J(Y)≤1}.It turns out that the optimal solution to the primal problem (2.1) can be obtained during the process of finding the constrained steepest ascend direction.

    ⅢThe modified augmented Lagrange multiplier (MALM) algorithm

    In this pant,we mention a mean-value technique for TMC problem[17].The problem can be expressed as the following convex programming,

    whereA,M∈Rn×nare both real Toeplitz matrices,? ?{?n+1,···,n?1}.LetD=P?(M).Then the partial augmented Lagrangian function is

    whereY∈Rn×n.

    Algorithm 2.2Step 0 Give?,sampled matrixD,μ0>0,ρ >1.Give also two initial matricesY0=0,E0=0.k:=0;

    Step 1 Compute the SVD of the matrix (D?Ek+) using the Lanczos method

    Step 2 Set

    Step 3 If‖D?Ak+1?Ek+1‖F(xiàn)/‖D‖F(xiàn)

    Step 4 SetYk+1=Yk+μk(D?Ak+1?Ek+1).Ifμk‖Ek+1?Ek‖F(xiàn)/‖D‖F(xiàn)

    RemarkIt is reported that MALM algorithm performs better that of much higher accuracy.Compared with the ALM,APGL,and SVT algorithms,the MALM algorithm is advantageous over the other three algorithms on the time costed by the SVD for smoothing at each iterate.

    As we know,the saving of the SVD time is at the expense of data communication.Sometimes,this is not worth the candle.This motivated us to put up with the following algorithm.

    ⅣThe?-step modified augmented Lagrange multiplier (?-MALM) algorithm

    To reduce the workload of data being moved at each iteration step,we propose a new accelerated algorithm for the TMC problem,which is smoothing once the diagonal elements of the iteration matrix by (1.5) for every?steps.The technique saves computation cost and reduces the data communication.It turns out that the iteration matrices keep a Toeplitz structure,which ensure the fast SVD of Toeplitz matrices can be utilized.

    Algorithm 2.3(?-MALM algorithm)

    Input:?,sampled matrixD,Y0,0=0,E0,0=0;parametersμ0>0,ρ >1,?,?1,?2.Letk:=0,q:=1,q=1,2,···,??1.

    Repeat:

    Step 1??1 iterations.

    1) Compute the SVD of the matrix (D?Ek,q+) using the Lanczos method

    2) Set

    3) If‖D?Xk+1,q+1?Ek+1,q+1‖F(xiàn)/‖D‖F(xiàn)

    4) SetYk+1,q+1=Yk,q+μk,q(D?Xk+1,q+1?Ek+1,q+1),μk+1,q+1=ρμk,q;otherwise,go to Step 1 1);

    Step 2?-th smoothing.

    1) Compute

    UpdateEk+1,?=

    Step 3 If‖D?Ak+1,??Ek+1,?‖F(xiàn)/‖D‖F(xiàn)

    Step 4 SetYk+1,q+1=Yk,q+μk,q(D?Ak+1,q+1?Ek+1,q+1).

    Ifμk,q‖Ek+1,q+1?Ek,q‖F(xiàn)/‖D‖F(xiàn)

    RemarkClearly,this algorithm is an acceleration of the MALM algorithm in [17].When?=1,it becomes the MALM scheme.

    3.Convergence Analysis

    We provided first some lemmas in the following.

    Lemma 3.1[6]LetA∈Rm×nbe an arbitrary matrix andUΣVTbe its SVD.Then the set of subgradients of the nuclear norm ofAis provided by

    ?‖A‖?={UVT+W:W∈Rm×n,UTW=0,WV=0,‖W‖2≤1}.

    Lemma 3.2[9]Ifμkis nondecreasing then each term of the following series is nonnegative and the series is convergent,that is,

    Lemma 3.3[9]The sequences{},{Yk}and{}are all bounded,where=Yk+1+μk?1(D?Ak?Ek?1).

    Lemma 3.4The sequence{Yk,q} generated by Algorithm 2.3 is bounded.

    ProofLetdefined as (1.5).

    First of all,we indicate thatYk,q,Ek,q,k=1,2,···,q=1,2,···,??1 are all Toeplitz matrices.Evidently,Y0,0=0,E0,0=0 are both smoothed into Toeplitz matrices.Suppose that afterYk,q,Ek,qare both Toeplitz matrices,so isEk+1,q+1=Thus,Yk+1,q+1is a Toeplitz matrix also from the Step 4 in Algorithm 2.3.

    And,

    It is clear that by Steps 1-2 in Algorithm 2.3,

    Hence we can obtain thatYk,q+μk,q(D?Ak+1,q+1?Ek,q)∈?‖Ak+1,q+1‖?from Lemmas 3.2 and 3.3.It is known that forAk+1,q+1=UΣVTby Lemma 3.1,

    We have also,

    Therefore,the following inequalities can be obtained:

    and

    It is clear that the sequence{Yk,q} is bounded.

    Theorem 3.1Suppose thatthen the sequence{Ak,q} converges to the solution of (2.5) whenμk,q→∞and

    ProofIt is true that

    since(Yk+1,q+1?Yk,q)=D?Ak+1,q+1?Ek+1,q+1and Lemma 3.4.Let (?,?) be the solution of (2.5).ThenAk+1,q+1,Yk+1,q+1,Ek+1,q+1,k=1,2,···,are all Toeplitz matrices from+=D.We prove first that where=Yk,q+μk,q(D?Ak+1,q+1?Ek,q),is the optimal solution to the dual problem (2.3).

    We obtain the following result through the same analysis,

    Then

    holds true.

    On the other hand,the following is true by Algorithm 2.3:

    Moreover,along the same idea of Theorem 2 in [9],it is obtained thatis the solution of(2.5).

    Theorem 3.2LetX=(xij)∈Rn×n,T(X)=()∈Rn×nbe the Toeplitz matrix derived fromX,introduced in (1.5).Then for all Toeplitz matrixY=(yij)∈Rn×n,

    ProofBy the definition ofT(X),we have=0,i,j=1,2,···,n.SinceYis a Toeplitz matrix,andyl=yij,l=i?j,i,j=1,2,···,n.Then

    Theorem 3.3In Algorithm 2.3,Ak,qis a Toeplitz matrix derived byXk,q.Then

    where?is the solution of (2.5).

    Proof

    4.Numerical Experiments

    In this section,some original numerical results of two algorithms(MALM,?-MALM)are presented for then×nmatrices with different ranks.We conducted numerical experiments on the same and modest workstation.By analyzing and comparing iteration numbers (IT),computing time in second (time(s)),deviation (error 1,error 2) and ratio which are defined in the following,we can see that the?-MALM algorithm proposed by this paper is far more effective than the MALM algorithm.

    In our experiments,M∈Rn×nrepresents the Toeplitz matrix.We select the sampling densityp=m/(2n?1),wheremis the number of the observed diagonal entries ofM,then 0≤m≤2n?1.With regard to the?-MALM algorithm,we set the parametersτ0=1/‖D‖2,δ=1.2172+,?1=10?9,?2=5×10?6and?=3 as a rule of thumb.The parameters of the MALM algorithm take the same as the?-MALM algorithm.

    The experimental results of two algorithms are shown in Tables 4.1-4.4.From the tables,two algorithms can successfully calculate the approximate solution of prescriptive stop condition for all the test matricesM.And our?-MALM algorithm in computing time is far less than that of the MALM algorithm.In particular,compared with the CPU of the MALM algorithm,we can find that the CPU of the?-MALM algorithm is reduced to 45%.The“ratio”in Table 4.5 can show this effectiveness.

    Table 4.1 Comparison between MALM and ?-MALM for p=0.6.

    Table 4.3 Comparison between MALM and ?-MALM for p=0.4.

    Table 4.4 Comparison between MALM and ?-MALM for p=0.3.

    Table 4.5 The values of ratio.

    5.Conclusion

    As is known to all,matrix completion is usually to recover a matrix from a subset of the elements of a matrix by taking advantage of low rank structure matrix interdependencies between the entries.It is well-known but NP-hard in general.In recent years,Toeplitz matrix completion has attracted widespread attention and TMC is one of the most important completion problems.In order to solve such problems,we put forward an?-step modified augmented Lagrange multiplier (?-MALM) algorithm based on the MALM algorithm,and corresponding with the theory of the convergence of the?-MALM algorithm are established.Theoretical analysis and numerical results have shown that the?-MALM algorithm is effective for solving TMC problem.The?-MALM algorithm overcomes the original ALM algorithm both singular value decomposition of tardy,and surmounts the property of the extra load of the MALM algorithm.The reason is that data communication congestion is far more expensive than computing.Compared with the CPU of the MALM algorithm,we can see that the CPU of our?-MALM algorithm is reduced to 45%.Therefore,?-MALM algorithm has better convergence rate for solving TMC problem than the MALM algorithm (tables 4.1-4.5).

    AcknowledgmentsThe authors gratefully acknowledge the anonymous referees and Professor ZZ Bai(academy of mathematics and systems science,Chinese academy of sciences)for their helpful comments and suggestions which greatly improved the original manuscript of this paper.

    中文字幕高清在线视频| 欧美日韩成人在线一区二区| 国产精品二区激情视频| 国产日韩欧美在线精品| 国产精品麻豆人妻色哟哟久久| 亚洲精品一二三| 99精国产麻豆久久婷婷| 午夜老司机福利片| 亚洲精品久久午夜乱码| 国产乱来视频区| 精品一区二区三卡| 最近最新中文字幕大全免费视频 | 又大又黄又爽视频免费| 啦啦啦在线免费观看视频4| 久久人人爽人人片av| 国产免费又黄又爽又色| 天堂中文最新版在线下载| 亚洲欧洲国产日韩| 久久99热这里只频精品6学生| a 毛片基地| 91成人精品电影| 欧美黑人欧美精品刺激| 中文欧美无线码| 国产激情久久老熟女| 高清不卡的av网站| 青春草国产在线视频| 秋霞伦理黄片| 国产精品久久久久久人妻精品电影 | 免费日韩欧美在线观看| 国产色婷婷99| 欧美日韩综合久久久久久| 国产熟女午夜一区二区三区| 国产成人免费无遮挡视频| 青春草国产在线视频| 赤兔流量卡办理| 成人毛片60女人毛片免费| 99久久精品国产亚洲精品| 日日摸夜夜添夜夜爱| 激情视频va一区二区三区| 久久久精品免费免费高清| 男女之事视频高清在线观看 | 中文精品一卡2卡3卡4更新| av卡一久久| 欧美成人精品欧美一级黄| 18禁国产床啪视频网站| 街头女战士在线观看网站| 多毛熟女@视频| 欧美国产精品va在线观看不卡| 97在线人人人人妻| 90打野战视频偷拍视频| 七月丁香在线播放| 精品少妇久久久久久888优播| 国产一区亚洲一区在线观看| 一边亲一边摸免费视频| 国产成人啪精品午夜网站| 亚洲美女视频黄频| 777米奇影视久久| 99热网站在线观看| 精品久久蜜臀av无| 99热网站在线观看| 一级毛片电影观看| 亚洲欧美激情在线| 亚洲,欧美,日韩| 日本vs欧美在线观看视频| 久久青草综合色| 另类精品久久| www.av在线官网国产| 久久99精品国语久久久| 精品少妇久久久久久888优播| 丰满饥渴人妻一区二区三| 97在线人人人人妻| videos熟女内射| 在线天堂中文资源库| 成人黄色视频免费在线看| 人人澡人人妻人| 丰满饥渴人妻一区二区三| 黄片小视频在线播放| 色精品久久人妻99蜜桃| 久久久久久久久免费视频了| 狠狠精品人妻久久久久久综合| 捣出白浆h1v1| 免费观看a级毛片全部| 日韩成人av中文字幕在线观看| 久久久久久免费高清国产稀缺| 亚洲成av片中文字幕在线观看| 亚洲精品久久久久久婷婷小说| 伦理电影免费视频| 久久精品久久久久久久性| 黄片小视频在线播放| 大香蕉久久成人网| 成人18禁高潮啪啪吃奶动态图| 色婷婷久久久亚洲欧美| 国产xxxxx性猛交| 亚洲欧美一区二区三区黑人| 国产片内射在线| 亚洲精品成人av观看孕妇| 亚洲精品一二三| 亚洲精品,欧美精品| 人人澡人人妻人| 黄色 视频免费看| 不卡视频在线观看欧美| 免费不卡黄色视频| 777久久人妻少妇嫩草av网站| 精品一品国产午夜福利视频| 亚洲精品,欧美精品| 观看美女的网站| 如何舔出高潮| 国产在线免费精品| 自拍欧美九色日韩亚洲蝌蚪91| 免费日韩欧美在线观看| 欧美老熟妇乱子伦牲交| 久久精品亚洲av国产电影网| 国产99久久九九免费精品| 中文字幕亚洲精品专区| 肉色欧美久久久久久久蜜桃| 又粗又硬又长又爽又黄的视频| 国产精品久久久久久精品电影小说| 国产精品久久久av美女十八| 亚洲精品日韩在线中文字幕| 国产成人av激情在线播放| 蜜桃国产av成人99| 叶爱在线成人免费视频播放| 国产精品三级大全| 晚上一个人看的免费电影| 黄色一级大片看看| xxx大片免费视频| 中文乱码字字幕精品一区二区三区| 日韩精品免费视频一区二区三区| 午夜av观看不卡| 久久97久久精品| 亚洲av成人不卡在线观看播放网 | 热99久久久久精品小说推荐| 成人手机av| 精品少妇久久久久久888优播| 三上悠亚av全集在线观看| 成人免费观看视频高清| 99久久精品国产亚洲精品| 免费在线观看视频国产中文字幕亚洲 | 王馨瑶露胸无遮挡在线观看| 精品人妻一区二区三区麻豆| 欧美老熟妇乱子伦牲交| 精品国产乱码久久久久久小说| 午夜日韩欧美国产| 国产精品久久久人人做人人爽| 亚洲精品日韩在线中文字幕| 国产免费一区二区三区四区乱码| av视频免费观看在线观看| 国产黄色免费在线视频| 在线观看免费高清a一片| 高清在线视频一区二区三区| 久久天躁狠狠躁夜夜2o2o | 啦啦啦啦在线视频资源| 精品国产乱码久久久久久男人| 美女扒开内裤让男人捅视频| 欧美日韩亚洲高清精品| 亚洲欧美一区二区三区黑人| 亚洲精品久久午夜乱码| 日韩视频在线欧美| 国产一区二区三区av在线| 卡戴珊不雅视频在线播放| 精品亚洲乱码少妇综合久久| 国产黄色免费在线视频| 国精品久久久久久国模美| 新久久久久国产一级毛片| 51午夜福利影视在线观看| www.精华液| 日韩一区二区三区影片| 老司机影院毛片| 大码成人一级视频| 蜜桃国产av成人99| 国产成人精品久久二区二区91 | 国产精品麻豆人妻色哟哟久久| 一级片免费观看大全| 国产精品女同一区二区软件| 新久久久久国产一级毛片| 国产女主播在线喷水免费视频网站| 国产精品一区二区在线不卡| 日本91视频免费播放| 人成视频在线观看免费观看| 日日爽夜夜爽网站| 少妇被粗大的猛进出69影院| 亚洲色图 男人天堂 中文字幕| 毛片一级片免费看久久久久| 亚洲精品久久成人aⅴ小说| 国产精品偷伦视频观看了| 国产成人免费无遮挡视频| 国产成人系列免费观看| 成人黄色视频免费在线看| 欧美乱码精品一区二区三区| avwww免费| 日本av免费视频播放| 亚洲婷婷狠狠爱综合网| 亚洲精华国产精华液的使用体验| 亚洲免费av在线视频| 国产老妇伦熟女老妇高清| 免费看不卡的av| 日韩伦理黄色片| 超碰成人久久| 久久久久精品人妻al黑| 一本色道久久久久久精品综合| 最近2019中文字幕mv第一页| 欧美精品人与动牲交sv欧美| 久久韩国三级中文字幕| 国产成人午夜福利电影在线观看| 99热网站在线观看| 秋霞在线观看毛片| 欧美最新免费一区二区三区| 久久久久人妻精品一区果冻| 男女午夜视频在线观看| 两性夫妻黄色片| 国产免费视频播放在线视频| 久久精品aⅴ一区二区三区四区| 一区在线观看完整版| 不卡av一区二区三区| 99国产精品免费福利视频| 国产一区亚洲一区在线观看| 亚洲图色成人| 一区在线观看完整版| 亚洲精品日本国产第一区| 精品少妇黑人巨大在线播放| 精品一区在线观看国产| 搡老岳熟女国产| 国产 一区精品| 欧美 亚洲 国产 日韩一| 国产伦理片在线播放av一区| 各种免费的搞黄视频| 国产无遮挡羞羞视频在线观看| 一级毛片黄色毛片免费观看视频| 啦啦啦在线免费观看视频4| 精品一品国产午夜福利视频| 国产精品亚洲av一区麻豆 | 别揉我奶头~嗯~啊~动态视频 | 亚洲综合精品二区| 美女高潮到喷水免费观看| 国产男女内射视频| 日韩制服丝袜自拍偷拍| 街头女战士在线观看网站| 日本欧美国产在线视频| 大陆偷拍与自拍| 两个人看的免费小视频| 成人午夜精彩视频在线观看| 亚洲av福利一区| 999久久久国产精品视频| av网站免费在线观看视频| 久久韩国三级中文字幕| 日本黄色日本黄色录像| 亚洲国产中文字幕在线视频| 91成人精品电影| 中文天堂在线官网| 在线观看国产h片| av网站在线播放免费| 久久久久久人人人人人| 亚洲成人av在线免费| e午夜精品久久久久久久| 美女国产高潮福利片在线看| 免费看不卡的av| 日韩大码丰满熟妇| 丁香六月欧美| 精品视频人人做人人爽| 大香蕉久久成人网| 亚洲精品乱久久久久久| 亚洲av在线观看美女高潮| videosex国产| 七月丁香在线播放| 美女国产高潮福利片在线看| 欧美黑人精品巨大| 亚洲综合色网址| a级毛片黄视频| 午夜福利一区二区在线看| 国产精品 国内视频| 亚洲婷婷狠狠爱综合网| 国产日韩欧美在线精品| 亚洲一区二区三区欧美精品| 大片免费播放器 马上看| 搡老岳熟女国产| 久久免费观看电影| 国产av国产精品国产| 两个人免费观看高清视频| 亚洲一区二区三区欧美精品| 人人妻人人添人人爽欧美一区卜| 黑人巨大精品欧美一区二区蜜桃| 国产精品国产av在线观看| 欧美精品一区二区大全| 亚洲自偷自拍图片 自拍| 国产色婷婷99| 国产精品国产三级专区第一集| 多毛熟女@视频| 无遮挡黄片免费观看| 国产成人精品久久久久久| 国产激情久久老熟女| 水蜜桃什么品种好| 国产福利在线免费观看视频| 国产精品 欧美亚洲| 丝袜喷水一区| 老鸭窝网址在线观看| 熟女少妇亚洲综合色aaa.| 2021少妇久久久久久久久久久| 国产成人欧美| 在线天堂中文资源库| 十分钟在线观看高清视频www| 在线 av 中文字幕| 日韩精品免费视频一区二区三区| 亚洲综合色网址| 久久精品久久久久久久性| 黄色一级大片看看| av片东京热男人的天堂| 男人爽女人下面视频在线观看| 少妇被粗大的猛进出69影院| 国产麻豆69| 18禁动态无遮挡网站| 精品国产一区二区三区久久久樱花| 亚洲精品,欧美精品| 久久久久国产一级毛片高清牌| av不卡在线播放| 国产av一区二区精品久久| 免费观看人在逋| 久久韩国三级中文字幕| 一区二区日韩欧美中文字幕| 一区二区av电影网| 国产一级毛片在线| 国产有黄有色有爽视频| 久久这里只有精品19| 亚洲国产欧美日韩在线播放| 国产男女超爽视频在线观看| 激情五月婷婷亚洲| 女的被弄到高潮叫床怎么办| 男女下面插进去视频免费观看| 在线观看国产h片| 老司机靠b影院| 街头女战士在线观看网站| 午夜日韩欧美国产| 人妻 亚洲 视频| 久久天躁狠狠躁夜夜2o2o | 日韩大片免费观看网站| 嫩草影视91久久| 满18在线观看网站| 啦啦啦在线免费观看视频4| 丝瓜视频免费看黄片| 日本色播在线视频| av不卡在线播放| 欧美人与善性xxx| 国产亚洲最大av| 18禁动态无遮挡网站| 欧美日韩综合久久久久久| 亚洲精品国产一区二区精华液| 在线观看一区二区三区激情| 街头女战士在线观看网站| h视频一区二区三区| 国产成人精品久久久久久| 日韩制服骚丝袜av| 久久精品亚洲熟妇少妇任你| 欧美变态另类bdsm刘玥| 观看av在线不卡| 国产成人欧美| 永久免费av网站大全| 国精品久久久久久国模美| 国产在线视频一区二区| 亚洲av福利一区| 久久久精品区二区三区| 青草久久国产| 2021少妇久久久久久久久久久| 成人午夜精彩视频在线观看| 天天躁夜夜躁狠狠久久av| 午夜福利视频精品| 久久天堂一区二区三区四区| 久久精品久久久久久噜噜老黄| 亚洲精品中文字幕在线视频| 成年美女黄网站色视频大全免费| 青青草视频在线视频观看| 一本一本久久a久久精品综合妖精| 日本wwww免费看| 9热在线视频观看99| 亚洲七黄色美女视频| 我的亚洲天堂| 日韩精品免费视频一区二区三区| 亚洲美女视频黄频| 熟妇人妻不卡中文字幕| 亚洲人成电影观看| 国产精品久久久久成人av| 国产av精品麻豆| 丰满迷人的少妇在线观看| 新久久久久国产一级毛片| 国产97色在线日韩免费| 亚洲精品在线美女| 美女大奶头黄色视频| 国产精品久久久久久精品电影小说| www.av在线官网国产| 日本猛色少妇xxxxx猛交久久| 超碰97精品在线观看| 亚洲天堂av无毛| 我要看黄色一级片免费的| 欧美变态另类bdsm刘玥| 9色porny在线观看| 久久狼人影院| 久久久久精品久久久久真实原创| 精品亚洲成a人片在线观看| 国产极品天堂在线| 飞空精品影院首页| 亚洲精品视频女| 少妇猛男粗大的猛烈进出视频| videos熟女内射| 精品久久蜜臀av无| 自线自在国产av| 国产福利在线免费观看视频| 97精品久久久久久久久久精品| 夫妻性生交免费视频一级片| 午夜日韩欧美国产| 日本黄色日本黄色录像| 男女边吃奶边做爰视频| 国产一卡二卡三卡精品 | 丰满少妇做爰视频| 欧美变态另类bdsm刘玥| 极品人妻少妇av视频| 亚洲成人一二三区av| 制服丝袜香蕉在线| 黑人欧美特级aaaaaa片| 久久精品国产亚洲av涩爱| 亚洲精品国产一区二区精华液| 狠狠精品人妻久久久久久综合| 极品人妻少妇av视频| 久久精品久久久久久久性| 最近最新中文字幕大全免费视频 | 如何舔出高潮| 天天操日日干夜夜撸| 中文字幕av电影在线播放| 80岁老熟妇乱子伦牲交| 久久久久久久久免费视频了| 精品一品国产午夜福利视频| 日韩制服骚丝袜av| 日韩制服丝袜自拍偷拍| 亚洲一卡2卡3卡4卡5卡精品中文| 韩国精品一区二区三区| 街头女战士在线观看网站| a级毛片黄视频| 一区福利在线观看| 青春草亚洲视频在线观看| 日韩大片免费观看网站| 最近的中文字幕免费完整| 免费高清在线观看日韩| 男女国产视频网站| 丁香六月天网| av网站在线播放免费| 天天躁夜夜躁狠狠躁躁| 91aial.com中文字幕在线观看| √禁漫天堂资源中文www| 婷婷成人精品国产| 国产女主播在线喷水免费视频网站| 99久久综合免费| 黄片小视频在线播放| 啦啦啦在线免费观看视频4| 免费看不卡的av| 久久久国产一区二区| 欧美老熟妇乱子伦牲交| √禁漫天堂资源中文www| 三上悠亚av全集在线观看| 成人国产av品久久久| 中文字幕精品免费在线观看视频| 纵有疾风起免费观看全集完整版| 亚洲精品成人av观看孕妇| 一区在线观看完整版| 亚洲av在线观看美女高潮| 久久久久久久久久久久大奶| 国产一级毛片在线| 最近最新中文字幕免费大全7| 美女午夜性视频免费| 久久久久精品久久久久真实原创| 搡老岳熟女国产| 黑人猛操日本美女一级片| 国产免费又黄又爽又色| 久久人人爽av亚洲精品天堂| 精品人妻在线不人妻| 搡老乐熟女国产| 精品人妻在线不人妻| 久久久久久久精品精品| 男女床上黄色一级片免费看| 亚洲欧美一区二区三区久久| 五月开心婷婷网| 久久国产亚洲av麻豆专区| 天天躁狠狠躁夜夜躁狠狠躁| 亚洲av在线观看美女高潮| 日韩视频在线欧美| 国产精品国产三级专区第一集| 午夜免费鲁丝| 亚洲图色成人| 精品国产乱码久久久久久男人| √禁漫天堂资源中文www| 亚洲精品在线美女| 波野结衣二区三区在线| 91精品国产国语对白视频| 伊人久久大香线蕉亚洲五| avwww免费| 婷婷色综合大香蕉| 亚洲精品av麻豆狂野| 国产又爽黄色视频| 麻豆av在线久日| 亚洲婷婷狠狠爱综合网| 波野结衣二区三区在线| 日本午夜av视频| 亚洲精品国产av成人精品| 秋霞伦理黄片| 最新的欧美精品一区二区| 亚洲av电影在线观看一区二区三区| 999精品在线视频| 国产黄色免费在线视频| 国产1区2区3区精品| 日韩 欧美 亚洲 中文字幕| 日韩大码丰满熟妇| 亚洲精品av麻豆狂野| 波多野结衣av一区二区av| 免费观看a级毛片全部| 性色av一级| 中文字幕人妻丝袜制服| 丰满乱子伦码专区| 黄色 视频免费看| 成人亚洲精品一区在线观看| 成人国语在线视频| 亚洲精品第二区| 免费在线观看完整版高清| 亚洲国产中文字幕在线视频| 国产无遮挡羞羞视频在线观看| 99精国产麻豆久久婷婷| 国产片内射在线| 在线亚洲精品国产二区图片欧美| 成人毛片60女人毛片免费| 多毛熟女@视频| 热re99久久精品国产66热6| 久久久久久久久久久免费av| 成人漫画全彩无遮挡| 亚洲av在线观看美女高潮| 久久天躁狠狠躁夜夜2o2o | 天天操日日干夜夜撸| 一本久久精品| 亚洲av日韩精品久久久久久密 | 捣出白浆h1v1| 国产老妇伦熟女老妇高清| 国产免费现黄频在线看| 侵犯人妻中文字幕一二三四区| 国产精品秋霞免费鲁丝片| 人人妻,人人澡人人爽秒播 | 午夜福利网站1000一区二区三区| 超碰97精品在线观看| 波多野结衣一区麻豆| 久久国产精品男人的天堂亚洲| 日韩中文字幕视频在线看片| 国产精品久久久人人做人人爽| 成人影院久久| 亚洲,欧美,日韩| 99久国产av精品国产电影| 男女之事视频高清在线观看 | 亚洲婷婷狠狠爱综合网| 一本一本久久a久久精品综合妖精| 无限看片的www在线观看| 精品国产乱码久久久久久男人| 在线观看www视频免费| 超色免费av| 如何舔出高潮| 少妇人妻久久综合中文| 亚洲精品第二区| 欧美日韩成人在线一区二区| 晚上一个人看的免费电影| 国产精品成人在线| 国产xxxxx性猛交| 亚洲国产欧美在线一区| 国产成人精品福利久久| 99久久99久久久精品蜜桃| 超色免费av| 婷婷成人精品国产| 91aial.com中文字幕在线观看| 久久久久久久国产电影| 国产亚洲午夜精品一区二区久久| 国产精品一区二区在线观看99| 一边摸一边做爽爽视频免费| 久久久久精品久久久久真实原创| 欧美日韩精品网址| 日韩不卡一区二区三区视频在线| 999久久久国产精品视频| 一级毛片电影观看| 欧美黄色片欧美黄色片| 久久久精品免费免费高清| 国产成人免费观看mmmm| 成年女人毛片免费观看观看9 | 国产在线免费精品| 国产亚洲欧美精品永久| 一区福利在线观看| 精品一区二区免费观看| 在线观看一区二区三区激情| bbb黄色大片| 国产精品嫩草影院av在线观看| 中文字幕人妻丝袜制服| 91精品国产国语对白视频| 色网站视频免费| 亚洲 欧美一区二区三区| 午夜免费观看性视频| 国产精品 欧美亚洲| 青春草视频在线免费观看| 夫妻午夜视频| 国产毛片在线视频| 丰满少妇做爰视频| 满18在线观看网站| 久久精品国产a三级三级三级| 七月丁香在线播放| 在线天堂中文资源库| 欧美日韩国产mv在线观看视频| 欧美日韩视频精品一区| 国产深夜福利视频在线观看| 久久影院123| 精品少妇黑人巨大在线播放| 国产精品一区二区精品视频观看| 亚洲伊人久久精品综合| 男人添女人高潮全过程视频| netflix在线观看网站| 国产熟女欧美一区二区| 国产在视频线精品| 天天操日日干夜夜撸| 久久人人爽人人片av|