TWI401905B - Multiple Input Multiple Output Detection Method and Detector - Google Patents

Description

Multiple input multiple output detection method and detector

The invention relates to a detector and a detection method thereof, in particular to a multi-input multi-output detector and a detection method thereof.

Multiple-input multiple-output (MIMO) technology has been widely incorporated into various wireless communication standards to improve spectral efficiency. In a multiple-input multiple-output system, the transmitting end cuts the data into a multi-symbol stream, and uses a plurality of antennas to respectively transmit the stream of characters at the same time and in the same frequency band; the receiving end receives the words by using a plurality of antennas. The sum of the stream flows and separates the stream of characters to obtain the characters originally transmitted by the transmitting end. It is currently recognized that the receiver based on the most likely panoramic detection (MLD) has the best performance (ie, the lowest error rate), but its computational complexity is very high. In 2002, Kim et al. proposed a combination of QR decomposition (QRD) and M algorithm (hereinafter referred to as QRD-M detection), which can be found through tree search. The most approximate approximation solution, while reducing the computational complexity, still retains the best performance that is most likely to be detected.

I. Signal model

In a MIMO system with N _t transmit antennas and N _r receive antennas, assuming that the channel between the transmit antenna and the receive antenna is a flat fading channel, the word transmitted by the transmit antenna The relationship between the element and the signal received by the receiving antenna can be expressed as: Or abbreviated as y=Hx+n , where x is a transfer character vector, x _j is the character transmitted by the jth transmit antenna, and is a constellation point on a constellation; y is a receive The signal vector, y _i is the signal received by the ith receiving antenna; H is a channel matrix, h _ij is the channel response between the jth transmitting antenna and the ith receiving antenna, and can be assumed to be independent and identical Independent and identically distributed complex Gaussian random variable having an average of 0 and a variance of 1; n is a noise matrix, n _i is the impurity received by the ith receiving antenna And can be assumed to be independent and identically distributed complex Gaussian random variables with an average of 0 and a variance of .

II. QR decomposition

By QR decomposition, the channel matrix H can be expressed as H=QR , where Q is a unitary matrix of size N _r ×N _t and having orthogonal rows (satisfying Q ^H Q=I , superscript ^H Represents a conjugate complex transpose operation; R is an upper triangular matrix of size N _t ×N _t and the diagonal elements are real numbers. Multiplying the conjugate complex transposed matrix Q ^H of a single matrix by the received signal vector y yields a processed received signal vector z=Q ^H y=Rx+w , or is detailed as: Where w=Q ^H n and has the same probability density function as the noise vector n (ie w ~CN(0, I )). Therefore, the most approximate solution can be expressed as: Among them, ∥. The trick is the Euclidean distance. According to this, a cost function can be defined:

III. Tree diagram

Referring to FIG. 1, if the constellation used by the MIMO system includes | C | constellation points, a tree diagram having N _t +1 levels (ie, depth N _t ) can be defined as follows: in the 0th hierarchy roots extending | C | article M. (Branch) to the next class, then for each arrival path (path) extending next repeated operation class, class until N _t. Therefore, the number of paths to the i-th level ( i = 1, ..., N _t ) is | C | ^t , and for each of these paths, the jth ( j =0,...,| C |-1) Branches extending to the next level represent possible transfer characters Is the jth constellation point c _j in the constellation diagram.

Each path to the i-th level has i- segment branches and represents a possible transport byte . Therefore, the entire tree diagram has a total Strips with N _t segment branches, each path representing a possible transfer character vector x , and this The strip paths exactly represent the transfer character vector x , respectively. Kind of possibility. For example, FIG. 1 is a tree diagram formed according to a multiple input multiple output system having three transmit antennas and three receive antennas and using BPSK modulation, wherein the thick line path represents a possible transfer character vector x is { x ₃ =-1, x ₂ =1, x ₁ =1}.

IV. Tree search

In the tree diagram, for any path to the i-th level ( i =1,..., N _t ), its state is defined as the transfer character group represented by this path. . For the root of the tree, its jth ( j =0,...,| C |-1) extends to the branch of the next level (representing the jth constellation point c _j in the constellation) Metric, BM) is defined as: among them, It is related to the j-th branch that extends from the root to the next. For any path that reaches the i - th level ( i = 2,..., N _t ), its jth ( j =0,..., | C |-1) extends to the next level of branching ( The branch length representing the jth constellation point c _j ) in the constellation diagram is defined as: among them, Related to the state in which the path reaches the i - th level, And this path is related from the i - th level to the j-th branch of the next level. The path metric (PM) of a path is defined as the sum of the branch lengths of the branches on this path. For example, for the thick line path in Figure 1, the length of each branch is | z ₃ - r ₃₃ (-1)| ² , | z ₂ - r ₂₃ (-1)- r ₂₂ (1)| ² And | z ₁ - r ₁₃ (-1)- r ₁₂ (1)- r ₁₁ (1)| ² , the path length is the sum of the lengths of these branches.

According to the above definition, the first term in the formula (4) corresponds to the branch length of the branch in the tree diagram extending to the next level, and the i-th term in the formula (4) ( i = 2, ..., N _t ) corresponds to the branch length of the branch that reaches the i - th level in the tree diagram and extends to the branch of the next level. Obviously, the path of any path with a N _t segment branch is the cost of the transfer character vector x represented by this path. Therefore, through the tree search to find the one with the shortest path length in the path with the N _t segment branch, the most approximate solution can be obtained. .

V.M algorithm

Because the above tree search needs to be calculated The path of the strip path is long. When the number of transmission antennas N _t and the number of constellation points included in the constellation diagram | C | increase, the tree diagram will become larger, and thus the branch length of the required calculation will increase, which will increase. Computational complexity. In order to reduce the computational complexity, by the M algorithm, in the i-th hierarchy ( i =1,..., N _t -1) in the tree diagram, only the optimal path of M _i is retained (ie, the path length of the M _i path is the shortest) path), each extend to | C | article branches to the next class, while the remaining paths deleted to reduce the branch length calculations required. These reserved paths are called survivor paths. The M _i of each level in the tree diagram can be adjusted according to the corresponding channel conditions and path length (this is called the adaptive M algorithm). When the channel conditions are good, the M _i can be lowered to reduce the calculation. the complexity.

Improvement of VI.QRD-M algorithm

With the QRD-M algorithm, when the M _i residual paths extend | C | branches to the next level, the branch length of the M _i | C | branches still needs to be calculated to get from M _i | C | The best path of M _{i +1} is reserved in the path. Higuchi et al. proposed an improved method in the IEEE GLOBECOM 2004 pages 2480~2486 "Adaptive Selection of Surviving Symbol Replica Candidates Based on Maximum Reliability in QRM-MLD for OFCDM MIMO Multiplexing", each of which extends in the M _i residual path | C | When branching to the next level, multiple quadrant detection is used to rank | C | branches, and the branch length of M _{i +1} branches is calculated accordingly (corresponding to M _i | C | M _{i +1} best path to be preserved in the path), which can reduce the branch length required for the QRD-M algorithm. However, multi-quad detection is only applicable to checkerboard constellations, such as the 16-QAM constellation. Therefore, this improved method is not suitable for multi-input multi-output systems using M-PSK (M>4) modulation.

Nagayama et al. proposed another improvement method in the IEEE VTC-2006 Fall "A Proposal of QRM-MLD for Reduced Complexity of MLD to detect MIMO signals in Fading Environment", which uses quadrant detection when the surviving path of the M _i is extended. Let each path extend only | C | _α (<| C |) branches to the next level, and calculate the branch length of the M _i | C | _α branches to get from the M _i | C | _α path The M _{i +1} best path is preserved, which reduces the branch length required for the QRD-M algorithm. However, | C | _α and M _i in this paper are fixed constants. If you want to achieve near-most probable performance, you must decide | C | _α and M _i based on the worst channel conditions. However, according to this decision | C | _α and M _i are usually not small. Therefore, the degree of reduction in computational complexity will be very limited.

Therefore, the object of the present invention is to provide a multi-input and multi-output detection method, which can be combined with various modulation techniques, and has low computational complexity.

Therefore, the MIMO detection method of the present invention comprises a tree search for multi-input and multi-output systems using QR decomposition, the multi-transmission The multi-input multi-output detector includes a channel estimator, a QR decomposition unit, a conversion unit and a detection unit, and the channel estimator is based on a Receiving a signal vector for channel estimation to obtain a channel matrix, the QR decomposition unit performing channel decomposition on the channel matrix to obtain a single matrix and an upper triangular matrix, the conversion unit according to the single matrix, the single matrix The conjugate complex transposed matrix is multiplied by the received signal vector to obtain a processed received signal vector, and the MIMO detection method is performed by the detecting unit, and includes the following steps: (A) according to the upper triangular matrix For each path arriving at a current level, the following sub-steps are performed: (A1) for the current level, based on the corresponding received signal in the processed received signal vector, a de-interference reception with the removal of the previous level interference is calculated. a signal; and (A2) for each of a subset of a first target number segmented by a constellation: finding the distance in the subset Scrambling nearest constellation point receives a signal, and a representative of the path extending from the branching point in the constellation to the next hierarchy, and calculates the path length of the branch, and a branching path corresponding long.

Another object of the present invention is to provide a MIMO multi-output detector that can be combined with a variety of modulation techniques and has a low computational complexity.

Therefore, the multiple input multiple output detector of the present invention is suitable for tree search in a multiple input multiple output system using QR decomposition, and includes a channel estimator, a QR decomposition unit, a conversion unit and a detection. unit. The pass The channel estimator performs channel estimation based on a received signal vector to obtain a channel matrix. The QR decomposition unit performs channel decomposition on the channel matrix to obtain a single matrix and an upper triangular matrix. The conversion unit multiplies the conjugate complex transposed matrix of the single matrix by the received signal vector according to the single matrix to obtain a processed received signal vector. The detecting unit performs, according to the upper triangular matrix, for each path that reaches a current level, for the current level, calculating the removed front layer according to the corresponding received signal in the processed received signal vector. Disturbing the interference to receive the signal; and performing, for each of a subset of the first target number divided by a constellation: finding a constellation point in the subset that is closest to the de-interfering received signal, and from the The path extension 1 represents the branch of the constellation point to the next level, and calculates the branch length of the branch and the path length of a corresponding path.

The foregoing and other technical aspects, features and advantages of the present invention will be apparent from the following description of the preferred embodiments.

Referring to FIG. 2, a preferred embodiment of the multiple input multiple output detector of the present invention is suitable for performing tree search in a QR input based multiple input multiple output system, and receiving N _t via N _r receiving antennas 25. A signal of an antenna (not shown) is transmitted to obtain a received signal vector y .

The multi-input multi-output detector of the embodiment comprises a channel estimator 21, a QR decomposition unit 22, a conversion unit 23 and a detection unit 24. The channel estimator 21 performs channel estimation based on the received signal vector y to obtain a channel matrix H. The QR decomposition unit 22 performs QR decomposition on the channel matrix H to obtain a single matrix Q and an upper triangular matrix R , and H = QR . The converting unit 23 multiplies the conjugate complex transposed matrix Q ^{H of the} single matrix by the received signal vector y according to the single matrix Q to obtain a processed received signal vector z , and therefore, z = Q ^H y , or is detailed as (2).

Referring to FIG. 3 and FIG. 4, the multiple input multiple output detection method used by the detecting unit 24 includes the following steps: Step 31 is to determine, for the 0th layer, a constellation including | C | constellation points according to corresponding channel conditions. One of the subsets into which the graph is divided into the first target number S ₁ (1 S ₁ | C |), and the constellation map is divided according to the first target number S ₁ (the detailed decision mode and the split mode are left to be explained after the entire method is introduced).

Step 32 is based on the upper triangular matrix R , and performs the following substeps for the root of the tree at the 0th level: substep 321 is to perform the following substeps for each of the S ₁ subsets divided in step 31: substep 3211 is to find the distance received signal in the subset The nearest constellation point (The detailed search method is left to be explained after the whole method is introduced).

Sub-step 3212 is to extend from the root of the tree to represent a constellation point Branching to the first level, and calculating the branch length of the branch and a corresponding path (representing a transfer character group) The path is long and the calculation is as follows:

Step 33 is to determine, for the 0th level, _one of the paths to be retained in the S1 path extending to the first level in step 32 according to the corresponding channel condition and path length, and the second target number M ₁ (1) M ₁ S ₁ ).

Step 34 is to delete the remaining paths by retaining M ₁ shortest paths for the S ₁ paths extending to the first level in step 32.

In step 33 and step 34, the existing plurality of adaptive M algorithms can be used to determine the second target number M ₁ and retain the M ₁ shortest path, which will not be described here.

Step 41 is to set i = 2.

Step 42 is to determine, for the i - th level (ie, the current level), the first target number S _i (1) of the subset into which the constellation is divided according to the corresponding channel condition. S _i | C |), and the constellation map is divided according to the first target number S _i (the detailed decision mode and the split mode are left to be explained after the entire method is introduced).

Step 43 is for each of the M _{i -1} paths to the i - th level (representing a transmission character group) Performing the following sub-steps: sub-step 431 is for the i - th level, according to the corresponding one of the processed received signal vectors z , calculating the interference cancellation received signal after removing the previous layer interference The calculation method is as follows:

Sub-step 432 is to perform the following sub-steps for each of the S _i subsets segmented in step 42: sub-step 4321 is to find the distance de-interference received signal in the subset The nearest constellation point (The detailed search method is left to be explained after the whole method is introduced).

Sub-step 4322 is to extend from the path to represent a constellation point Branch to the i-th level (ie, the next level), and calculate the branch length of the branch and a corresponding path (representing a transfer character group) The path is long and the calculation is as follows:

Step 44 is to determine, for the i - th level, the second target number of paths to be retained in the M _{i -1} S _i path extending to the i-th level in step 43 according to the corresponding channel condition and path length. M _i (1 M _i M _{i -1} S _i ).

Step 45 is to delete the remaining paths by retaining the M _i shortest paths for the M _{i -1} S _i paths extending to the i-th level in step 43.

In step 44 and step 45, the existing various adaptive M algorithms can be used to determine the second target number M _i and the reserved M _i shortest path, which will not be described here.

Step 46 is to determine whether i is equal to N _t (ie, whether the next level is a final level), if not, skip to step 47, and if yes, go to step 48.

Step 47 is to increment i by 1, and skip to step 42.

Step 48 from reaching the first class N _t Select a shortest path from the path, and the transfer character group represented by the path is the detected transfer character vector. .

Since it is only necessary to calculate the branch length of the S ₁ branch in step 32, it is only necessary to calculate the branch length of the branch of M _{i -1} S _i ( i = 2, ..., N _t ) in step 43, therefore, S ₁ ~ When at least one of them is less than | C |, the branch length required for the QRD-M algorithm can be reduced. Moreover, S ₁ ~ And M ₁ ~ Appropriate adjustments based on channel conditions, while significantly reducing computational complexity, still achieve near-most probable detection performance.

Segmentation of constellation

In steps 31 and 42, the constellation map is segmented into S _i sub-sets using Ungerboeck's set partitioning (the first target number S _i must be a power of 2). For example, referring to FIG. 5, the 16-QAM constellation diagram can be divided into a subset A ₀ including sixteen constellation points, two sub-sets B ₀ and B ₁ each including eight constellation points, and each of which includes four Four sub-sets C ₀ ~ C _{3 of} constellation points and repulsive, eight sub-sets D ₀ ~ D ₇ each including two constellation points and repulsive, or sixteen sub-sets each including one constellation point and repelling (eg E ₀ in FIG. 5 ~ E ₁₅ shown); see FIGS. 6,8-PSK constellation may be divided into eight comprise a subset of the constellation point F _0, and each comprise four constellation points of two subsets repulsive G ₀ and G ₁ , four sub-sets H ₀ to H ₃ each including two constellation points and being repulsive, or eight sub-sets I ₀ to I ₇ each including one constellation point and being repulsive. Wenger's set partitioning makes the intraset distance of the subsets the largest among all the partitions, which can reduce error propagation.

Recent search for constellation points

In sub-step 3211 and sub-step 4321, when S _i <| C | (ie, the number of constellation points included in the subset is greater than 1), at least one threshold TH is used to cut the majority in the subset to respectively cover the constellation Point decision interval and judge the received signal And de-interference receiving signals (Where collectively referred to as signal z ) is the decision interval in which to locate the constellation point closest to the distance signal z .

Refer to Figure 5 for an example. When S _i = 4, the 16-QAM constellation can be divided into four sub-sets C ₀ ~ C ₃ each including four constellation points, as shown in Table 1:

The critical value TH of each sub-set C ₀ ~ C ₃ is shown in Table 2:

among them, , Is the minimum distance between constellation points of the received constellation, d _min is the minimum distance between constellation points of the constellation transmitted. Each threshold TH cuts four decision intervals in the corresponding subset C ₀ ~ C ₃ , wherein each decision interval covers a constellation point. The judgment rules are shown in Table 3: Table 3

The signal z is as shown by × in FIG. Since in the subset C ₀ , the signal z satisfies z >Re(TH) and satisfies z >Im(TH), the constellation point found is c ₂ ; since in the subset C ₁ , the signal z does not satisfy z >Re(TH), but satisfies z >Im(TH), so the constellation point found is c ₁₀ ; since in the sub-set C ₂ , the signal z does not satisfy z >Re(TH), but satisfies z >Im ( TH), so the constellation point found is c ₁₁ ; since the signal z satisfies z >Re(TH) and satisfies z >Im(TH) in the subset C ₃ , the constellation point found is c ₃ .

Referring to Figure 6, another example is illustrated. When S _i = 2, the 8-PSK constellation can be divided into two subsets G ₀ , G ₁ each including four constellation points, wherein the subset G ₀ can easily find the distance from a threshold TH The nearest constellation point of the signal z , and after the sub-set G ₁ is rotated by 45°, it is also possible to easily find the constellation point closest to the distance signal z with a critical value TH.

It should be noted that this embodiment can also be combined with other QAM modulation techniques, other PSK modulation techniques, and other modulation techniques, not limited to 16-QAM and 8-PSK.

First target number S _i decision

It is assumed that in the process of selecting branches, the correct path (representing the correct transfer character vector X ^true ) is not deleted. When this correct path is to be extended from the i - th level ( i = 1, ..., N _t ) to the next level, the signal z is as follows:

among them, Is the correct transfer character, It is a noise and is a complex Gaussian function.

Referring to Figure 7, in the case where the noise power is fixed (as shown by the circle in Figure 7), if the channel attenuation is severe (ie When it is small, the distance between the constellation points of the received constellation (shown by the circular points in Figure 7) will become smaller, and the noise will be small. May cause signal points Crossing a few constellation points and shifting to the signal z (as indicated by the diamond dots in Fig. 7), therefore, it is necessary to adopt a larger first target number S _i so that the found distance signal z is the closest S _i Constellation points can cover signal points . As shown in Fig. 7(a), if the channel attenuation is more serious, the circle covers more constellation points (that is, the first target number S _{i is} larger), as shown in Fig. 7(b), if the channel attenuation is slight, The circle covers fewer constellation points (ie, the first target number S _{i is} smaller). As can be seen from the above, the first target number S _{i is} related to the channel conditions (ie, channel attenuation and noise power). In the tree search, for the i - th level ( i = 1, ..., N _t ), the appropriate first target number S _i should be determined according to the corresponding channel condition.

When extending to the ith level, it is generally desirable to determine an appropriate first target number S _i such that the probability that the correct branch is deleted P _miss ( S _i ) is less than a target probability P _target . Obviously, the larger the first target number S _i , the smaller the P _miss ( S _i ), but the higher the computational complexity. Therefore, it is necessary to find a minimum first target number S _i that satisfies the above conditions to reduce computational complexity.

P _miss ( S _i ) is equivalent to the probability of event E : the branch length of the correctly transmitted character is not within the shortest branch length of the S _i . Since the calculation process of finding the solution of P _miss ( S _i )< P _target according to the probability of the event E is very complicated, it is necessary to calculate | C | branch lengths to be obtained. In this embodiment, the approximate solution is obtained by the following method. To reduce the computational complexity.

Referring to Figure 8, in the constellation diagram, a square covering the S _i constellation points is defined centering on the signal z . Event E can be approximated as event A : the correct transfer character falls outside the square. According to Forney's continuous approximation, the side length of a square can be approximated . According to formula (12) and ~CN(0, ), the probability that the correct branch is deleted P _miss ( S _i ) is as follows:

among them, . Solving P _miss ( S _i )< P _target gives a desired number :

Where Q ^-1 (.) is the inverse of Q (.), and S _max is the maximum number of first targets S _i that the hardware and software can tolerate. If S _{max is} not limited by software and hardware, S _max can be set to the number of constellation points included in the constellation | C |, for example, for a 16-QAM constellation, S _max can be set to 16.

In order to be able to utilize Wenger's set partitioning, the first target number S _i must be a power of 2, so ,among them, Is the nearest power but not less than the power of 2 of x .

In summary, this embodiment is in S ₁ ~ When at least one of them is less than | C |, the branch length calculated by the QRD-M algorithm can be reduced, and the embodiment can cooperate with various modulation techniques such as QAM and PSK, so that the object of the present invention can be achieved.

However, the above is only the preferred embodiment of the present invention, when not The scope of the invention is to be construed as being limited by the scope of the invention and the scope of the invention.

21‧‧‧channel estimator

22‧‧‧QR decomposition unit

23‧‧‧Transition unit

24‧‧‧Detection unit

25‧‧‧ receiving antenna

31~48‧‧‧Steps

1 is a tree diagram illustrating a tree search in a multiple input multiple output system; FIG. 2 is a block diagram illustrating a preferred embodiment of the multiple input multiple output detector of the present invention; FIGS. 3 and 4 are flow charts The detection method used in the preferred embodiment is illustrated; FIG. 5 is a schematic diagram showing how the set partitioning of Wengebo is applied to the 16-QAM constellation diagram; FIG. 6 is a schematic diagram showing how the Wengebo collection segmentation is applied. In the 8-PSK constellation diagram; Fig. 7 is a schematic diagram showing the relationship between the number of constellation points required and the channel conditions; and Fig. 8 is a schematic diagram showing how the number of constellation points required in this embodiment is determined.

41~48‧‧‧Steps

Claims (19)

A multiple input multiple output detection method is suitable for tree search in a multiple input multiple output system using QR decomposition, the multiple input multiple output system comprising a multiple input multiple output detector, the multiple input multiple output detection The detector includes a channel estimator, a QR decomposition unit, a conversion unit and a detection unit, and the channel estimator performs channel estimation according to a received signal vector to obtain a channel matrix, and the QR decomposition unit The channel matrix performs channel decomposition to obtain a single matrix and an upper triangular matrix, and the conversion unit multiplies the conjugate complex transposed matrix of the single matrix by the received signal vector according to the single matrix to obtain a processed received signal. The vector, the multiple input multiple output detection method is performed by the detecting unit, and includes the following steps: (A) according to the upper triangular matrix, for each path reaching a current level, performing the following substeps: (A1) The current hierarchy calculates a de-interfering received signal after removing the previous layer interference according to the corresponding received signal in the processed received signal vector; and (A2) performing, for each of a subset of the first target number divided by a constellation: finding a constellation point in the subset that is closest to the de-interfering received signal, and extending from the path to represent the The branch of the constellation point branches to the next level, and the branch length of the branch and the path length of a corresponding path are calculated. According to the multi-input and multi-output detection method described in claim 1 of the patent application scope, the following steps are further included: (B) If the next level is not a final level, move to the next level and jump to step (A). According to the multiple input multiple output detection method described in claim 1, the method further comprises the following steps: (C) retaining a second target number for all paths extending to the next level in step (A) The shortest path and the remaining paths are deleted. According to the multiple input multiple output detection method described in claim 3, the method further includes the following steps: (D) determining, for the current hierarchy, the second target number according to the corresponding channel condition and the path length. According to the multiple input multiple output detection method described in claim 4, in step (C) and step (D), an adaptive M algorithm is used to determine the second target number and retain the second target. The shortest path to the number. According to the multiple input multiple output detection method described in claim 1, the method further includes the following steps: (E) determining, for the current hierarchy, the first target number according to the corresponding channel condition, and according to the first The number of targets divides the constellation. According to the multiple input multiple output detection method described in claim 6 of the patent application, in step (E), the constellation map is segmented into a subset of the first target number by using Wengbo's set partitioning. According to the multiple input multiple output detection method described in claim 6 of the patent application, in step (E), a solution is obtained by finding a solution whose probability of occurrence is less than a target probability, and according to the Expected number gets this A first target number, wherein the event means that a correct transfer character falls outside a square that encompasses the first target number of constellation points in the constellation. According to the multiple input multiple output detection method described in claim 8 of the patent application, in the step (E), the first target number is the nearest power but not less than the power of 2 of the desired number. According to the multiple input multiple output detection method of claim 1, in sub-step (A2), when the number of constellation points included in the subset is greater than 1, at least one critical value is used in the subset A majority of the decision intervals respectively covering the constellation points are cut out, and it is determined in which decision interval the de-interference received signal is located to find the constellation point in the subset that is closest to the de-interfering received signal. A multi-input multi-output detector for tree search in a multiple input multiple output system using QR decomposition, and comprising: a channel estimator for channel estimation based on a received signal vector to obtain a a channel matrix; a QR decomposition unit that performs channel decomposition on the channel matrix to obtain a single matrix and an upper triangular matrix; a conversion unit that multiplies the conjugate complex transposed matrix of the single matrix by the single matrix Receiving a signal vector to obtain a processed received signal vector; and a detecting unit according to the upper triangular matrix, for each path reaching a current level, performing the following actions: for the current level, according to the processed receiving Signal direction a corresponding received signal in the quantity, calculating a de-interfering received signal after removing the previous layer interference; and performing a process for each of the subset of the first target number divided by a constellation: finding the distance in the subset The de-interference interferes with the nearest constellation point of the received signal, and extends from the path to a branch representing the constellation point to the next level, and calculates the branch length of the branch and the path length of a corresponding path. According to the multi-input multi-output detector of claim 11, wherein the detecting unit further retains a shortest path of a second target number and deletes the remaining paths for all paths extending to the next level. The multiple input multiple output detector according to claim 12, wherein the detecting unit determines the second target number for the current hierarchy according to the corresponding channel condition and the path length. The multiple input multiple output detector according to claim 13 , wherein the detecting unit is an shortest path that uses an adaptive M algorithm to determine the second target number and retain the second target number. According to the multi-input multi-output detector of claim 11, wherein the detecting unit determines the first target number according to the corresponding channel condition for the current hierarchical layer, and according to the first target The number divides the constellation. The multi-input multi-output detector according to claim 15 , wherein the detecting unit divides the constellation into a subset of the first target number by using Wengbo's set partitioning. According to the multi-input multi-output detector described in claim 15 of the patent application, The detecting unit obtains a desired number by finding a solution that the probability of an event is less than a target probability, and obtains the first target number according to the expected number, and the event refers to a correct transmission character. Outside of a square, the square encompasses the first target number of constellation points in the constellation. The multiple input multiple output detector of claim 17, wherein the first target number is the nearest power but not less than the desired number of powers of two. The multiple input multiple output detector according to claim 11, wherein the detecting unit uses at least one critical value in the subset when the number of constellation points included in the subset is greater than one A plurality of decision intervals respectively covering the constellation points are cut out, and it is determined in which decision interval the de-interference received signal is located to find the constellation point in the subset that is closest to the de-interfering received signal.