Enhanced digital predistortion

Abstract

In one embodiment of the present invention, a weight vector is generated based on a pre-distorted input signal and an output signal of an amplifier using a computation reduction technique. The computation reduction technique decomposes a number of multiplication operations between complex numbers such that a number of multiplications and a number of additions to generate the weight vector is reduced as compared to if the number of multiplication operations between complex numbers was not decomposed. An input signal for input to an amplifier is pre-distorted based on the generated weight vector.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

The present invention will become more fully understood from the detailed description given herein below and the accompanying drawings, wherein like elements are represented by like reference numerals, which are given by way of illustration only and thus are not limiting of the present invention and wherein:

FIG. 1 illustrates the prior art architecture for digital pre-distortion to correct amplifier distortion.

FIG. 2 illustrates the scaled relationship of the 5 elements in vector x(n), where x(n)=0.2+i0.8.

FIG. 3 graphically demonstrates that most of the elements in matrix R(n) are the same elements in matrix R(n−1).

DETAILED DESCRIPTION OF EXAMPLE EMBODIMENTS

The present invention provides methodologies for reducing the computation complexity of the pre-distorting operation discussed in detail above. To better understand the methodologies of the present invention, the operations for complex and scalar multiplications will be reviewed, and then the methodologies of the present invention will be described by examining the correlation matrix and the cross-correlation vector.

Complex and Scalar Multiplications

The basic operation to construct a correlation matrix is repeatedly multiplying two complex numbers and adding the result to accumulators. With the direct approach, each complex-complex multiplication requires 4 multiplications and 2 additions. An example of this is shown below:

(a+ib)·(c+id)=(a·c−b·d)+i(a·c+b·d)

Using this example, the distortion model will be analyzed and the scalar terms in the correlation matrix and cross-correlation vector identified. The general complex-complex multiplications will then be converted to scalar-complex multiplications or scalar-scalar multiplications according to methodologies of the present invention to reduce the computation complexity. The following are some guidelines:

• Sum of squares—Whenever a magnitude square is encountered, the sum-of-squares operation will be used, since this (e.g., |a+ib|²=a·a+b·b) requires only 2 multiplications and 1 addition.
• Scalar-complex multiplication—If one of the numbers is a scalar, scalar-complex multiplication will be used, since this (e.g., (a+ib)·c=a·c+ib·c) requires only 2 multiplications.
• Scalar-scalar multiplication—If both numbers are scalar, scalar-scalar multiplication will be used, since this requires only 1 multiplication.

Table 4 below summarizes the complexity of different multiplications:

TABLE 4
Multiplication
Operations	Equations	#Mult	# Add
Complex-complex	(a + ib) · (c + id) = (a · c −	4	2
	b · d) + i(a · c + b · d)
Sum of squares	|a + ib|2 = a · a + b · b	2	1
Scalar-complex	(a + ib) · c = a · c + ib · c	2	0
Scalar-scalar	a · b	1	0

Correlation Matrix

Reducing the computation complexity for determining the correlation matrix will be discussed first. The correlation matrix includes two types of blocks: Type 1 or diagonal blocks, and Type 2 or off-diagonal blocks.

Type 1 (Diagonal) Blocks

The correlation matrix has L=6 diagonal blocks. Without loss of generality, handling the first diagonal block R_1,1(n) will be described in detail and handling of the other diagonal blocks will be readily understood from this discussion. It will be shown that the K·K=25 elements in diagonal block R_1,1(n) are real-valued and can be represented by 9 polynomial terms.

Letting a=|x(n)|, we have:

$\underset{\_}{x}\left(n\right)=\left[\begin{array}{c}1\\ a\\ a^2\\ a^3\\ a^4\end{array}\right]x\left(n\right).$

Diagonal block R_1,1(n) is the auto-correlation block of x(n) and includes real valued elements as shown below:

$\begin{array}{c}{\underset{\_}{R}}_{1,1}\left(n\right)=\underset{\_}{x}\left(n\right){\underset{\_}{x}}^H\left(n\right)\\ =\left[\begin{array}{c}1\\ a\\ a^2\\ a^3\\ a^4\end{array}\right]\left[1aa^2a^3a^4\right]x\left(n\right)·x^{*}\left(n\right)\\ =\left[\begin{array}{ccccc}{a}^2& a^3& a^4& a^5& a^6\\ a^3& a^4& a^5& a^6& a^7\\ a^4& a^5& a^6& a^7& a^8\\ a^5& a^6& a^7& a^8& a^9\\ a^6& a^7& a^8& a^9& a^{10}\end{array}\right]\end{array}$

Notice that a²=|x(n)|²=x(n)·x*(n).

Therefore, there are only 9 distinct real values {a², a³, a⁴, . . . , a⁹ and a¹⁰} in the block. These polynomial terms can be obtained sequentially. Using these properties the computation complexity may be reduced.

The following is the procedure performed at the weight estimator 12 according to an embodiment of the present invention to compute the elements in diagonal block R_1,1(n):

• 1. Compute a², using sum-of-squares operation, where a²=|x(n)|²=(real(x(n)))²+(imag(x(n)))².
• 2. Compute a=sqrt(a²) and store value a along with x(n).
• 3. Sequentially compute a³, a⁴, . . . , a⁹ and a¹⁰, by multiplying the previous number by a, using scalar-scalar multiplications.

The computation complexity, excluding the square-root operation, is summarized below in table 5. The baseline given in this and other tables is for the prior art methodology discussed in detail in the Background of the Invention section:

	TABLE 5
	Operations	#Mult	#Add
	a2, Sum of squares	2	1
	a3, a4, . . . , a9 and a10	8	0
	Total	10	1
	Baseline: 15 complex-complex	60	30
	multiplications

Type 2 (Off-diagonal) Blocks

Without loss of generality, handling of off-diagonal block R_1,2(n) will be described in detail, and the handling of other off-diagonal blocks will be readily apparent from this discussion. Letting a=|x(n)| and b=|x(n−1)|, we have:

$\underset{\_}{x}\left(n\right)=\left[\begin{array}{c}1\\ a\\ a^2\\ a^3\\ a^4\end{array}\right]x\left(n\right)\mathrm{and}\underset{\_}{x}\left(n-1\right)=\left[\begin{array}{c}1\\ b\\ b^2\\ b^3\\ b^4\end{array}\right]x\left(n-1\right).$

The off-diagonal block becomes:

$\begin{array}{c}{\underset{\_}{R}}_{1,2}\left(n\right)=\underset{\_}{x}\left(n\right){\underset{\_}{x}}^H\left(n-1\right)\\ =\left[\begin{array}{ccccc}1& b& b^2& b^3& b^4\\ a& \mathrm{ab}& {\mathrm{ab}}^2& {\mathrm{ab}}^3& {\mathrm{ab}}^4\\ a^2& a^2b& a^2b^2& a^2b^3& a^2b^4\\ a^3& a^3b& a^3b^2& a^3b^3& a^3b^4\\ a^4& a^4b& a^4b^2& a^4b^3& a^4b^4\end{array}\right]x\left(n\right)·x^{*}\left(n-1\right)\end{array}$

Therefore, the only complex-complex multiplication that is computed is the first element x(n)·x*(n−1) in the block. The rest of the terms may be obtained sequentially using scalar-complex multiplications.

The following is the procedure the weight estimator 12 performs according to an embodiment of the present invention to compute elements in the off-diagonal block R_1,2(n):

• 1. Compute the first element of R_1,2(n), x(n)·x*(n−1), using a complex-complex multiplication.
• 2. Compute the remaining elements (or terms) in the first row sequentially, by repeatedly multiplying the previous element by b, using scalar-complex multiplications.
• 3. Compute the remaining rows sequentially, by repeatedly multiplying the previous row by a, using scalar-complex multiplications.

The computation complexity is summarized below in table 6 below:

	TABLE 6
	Operations	#Mult	#Add
	1 complex-complex multiplication	4	2
	24 scalar-complex multiplications	48	0
	Total	52	2
	Baseline: 25 complex-complex	100	50
	multiplications

The magnitude b=|x(n−1)| may be computed and stored at the weight estimator 12 while computing diagonal block R_1,1(n−1) at time n−1. Accordingly, there is no need to repeat the square-root operation.

The following tables summarize the reduction in computation complexity for the improved algorithm. If we compute the 465 elements in the upper triangular correlation matrix according to the symmetry property, the complexity reduction is summarized in the following table:

	TABLE 7
	Operations	#Mult	#Add
	6 diagonal blocks	60	6
	15 off-diagonal blocks	780	30
	Total	840	36
	Baseline: 465 complex-complex	1860	930
	multiplications

If we compute only the 140 elements in the top block row of the correlation matrix according to the delayed version property, the complexity reduction is summarized in the following table:

	TABLE 8
	Operations	#Mult	#Add
	1 diagonal block	10	1
	5 off-diagonal blocks	260	10
	Total	270	12
	Baseline: 140 complex-complex	560	280
	multiplications

In both cases, more than one half of the computation complexity can be reduced.

Cross Correlation Vector

Next, reducing the computation complexity for determining the cross correlation vector will be described.

The cross correlation vector at time n is:

$\begin{array}{c}\underset{\_}{v}\left(n\right)=d\left(n\right)·\underset{\_}{u}\left(n\right)\\ =\left[\begin{array}{c}d\left(n\right)·\underset{\_}{x}\left(n\right)\\ d\left(n\right)·\underset{\_}{x}\left(n-1\right)\\ d\left(n\right)·\underset{\_}{x}\left(n-2\right)\\ d\left(n\right)·\underset{\_}{x}\left(n-3\right)\\ d\left(n\right)·\underset{\_}{x}\left(n-4\right)\\ d\left(n\right)·\underset{\_}{x}\left(n-5\right)\end{array}\right].\end{array}$

Let a₁=|x(n−l)|, 0≦l≦5, then each cross correlation vector block d(n)·x(n−l) may be expressed as:

$d\left(n\right)·\underset{\_}{x}\left(n-l\right)=\left[\begin{array}{c}1\\ a_l\\ a_l^2\\ a_l^3\\ a_l^4\end{array}\right]d\left(n\right)·x\left(n-l\right),0\le l\le 5.$

Within this block, complex-complex multiplication is used once to compute d(n)·x(n−l), and the remaining 4 elements are obtained, sequentially, by multiplying the previous element by a₁, using scalar-complex multiplication. Note that a₁, 0≦l≦5 may be computed and stored before and does not need to be computed again.

The following is the procedure performed by the weight estimator 12 according to an example embodiment of the present invention to compute the cross-correlation vector:

Repeating for 0≦l≦5,

• 1. Compute v_5l+1(n)=d(n)·x(n−l), using complex-complex multiplication.
• 2. Compute v_5l+2(n), . . . , v_5l+5(n), sequentially, by multiplying the previous element by a_l, using scalar-complex multiplication.

The computation complexity is summarized below in table 7:

	TABLE 9
	Operations	#Mult	#Add
	6 complex-complex multiplications	24	12
	24 scalar-complex multiplications	48	0
	Total	72	12
	Baseline: 30 complex-complex	120	60
	multiplications

Improved Method with Cross Terms

Next, cross-polynomial terms will be considered, and it will be shown that the same techniques of the present invention may be used to reduce computation complexity. Cross-polynomial terms account for correlation between polynomial terms from different delay times, and may improve the effectiveness of the distortion modeling.

Consider the products |x(n−m)|²x(n) and |x(n−m)|⁴x(n), 1≦m≦12, which are the cross product between a current sample and a magnitude-polynomial at a different sample time. Let b_m²=|x(n−m)|² and b_m⁴=|x(n−m)|⁴. Two cross term vectors y(n) and z(n) may be defined as:

$\underset{\_}{y}\left(n\right)=\left[\begin{array}{c}{x\left(n-1\right)}^2\\ {x\left(n-2\right)}^2\\ {x\left(n-3\right)}^2\\ ⋮\\ {x\left(n-12\right)}^2\end{array}\right]x\left(n\right)=\left[\begin{array}{c}{b}_1^2\\ b_2^2\\ b_3^2\\ \cdots \\ b_{12}^2\end{array}\right]x\left(n\right)$ $\underset{\_}{z}\left(n\right)=\left[\begin{array}{c}{x\left(n-1\right)}^4\\ {x\left(n-2\right)}^4\\ {x\left(n-3\right)}^4\\ ⋮\\ {x\left(n-12\right)}^4\end{array}\right]x\left(n\right)=\left[\begin{array}{c}{b}_1^4\\ b_2^4\\ b_3^4\\ \cdots \\ b_{12}^4\end{array}\right]x\left(n\right)$

The cross term vectors y(n) and z(n) are appended to the concatenated amplifier output signal vector u(n) as shown below:

$\underset{\_}{u}\left(n\right)=\left[\begin{array}{c}\underset{\_}{x}\left(n\right)\\ \underset{\_}{x}\left(n-1\right)\\ \underset{\_}{x}\left(n-2\right)\\ \underset{\_}{x}\left(n-3\right)\\ \underset{\_}{x}\left(n-4\right)\\ \underset{\_}{x}\left(n-5\right)\\ \underset{\_}{y}\left(n\right)\\ \underset{\_}{z}\left(n\right)\end{array}\right]=\left[\begin{array}{c}{u}_1\left(n\right)\\ u_2\left(n\right)\\ u_3\left(n\right)\\ u_4\left(n\right)\\ ⋮\\ u_{53}\left(n\right)\\ u_{54}\left(n\right)\end{array}\right].$

Note that with the cross terms, there are M=KL+12+12=54 elements in the input signal vector u(n). This assumes that b_m² and b_m⁴ were computed and stored by the weight estimator 12 while computing diagonal block R_1,1(n−m) at time n−m for m=1, 2, . . . , 12. It is also assumed that a²=|x(n)|² was already computed and stored for use by the weight estimator 12.

Correlation Matrix

Reducing the computation complexity for determining the correlation matrix including cross polynomial terms will now be described. The expanded correlation matrix becomes:

$\underset{\_}{R}\left(n\right)=\left[\begin{array}{cccccccc}{\underset{\_}{R}}_{1,1}\left(n\right)& {\underset{\_}{R}}_{1,2}\left(n\right)& {\underset{\_}{R}}_{1,3}\left(n\right)& {\underset{\_}{R}}_{1,4}\left(n\right)& {\underset{\_}{R}}_{1,5}\left(n\right)& {\underset{\_}{R}}_{1,6}\left(n\right)& {\underset{\_}{R}}_{1,7}\left(n\right)& {\underset{\_}{R}}_{1,8}\left(n\right)\\ {\underset{\_}{R}}_{2,1}\left(n\right)& {\underset{\_}{R}}_{2,2}\left(n\right)& {\underset{\_}{R}}_{2,3}\left(n\right)& {\underset{\_}{R}}_{2,4}\left(n\right)& {\underset{\_}{R}}_{2,5}\left(n\right)& {\underset{\_}{R}}_{2,6}\left(n\right)& {\underset{\_}{R}}_{2,7}\left(n\right)& {\underset{\_}{R}}_{2,8}\left(n\right)\\ {\underset{\_}{R}}_{3,1}\left(n\right)& {\underset{\_}{R}}_{3,2}\left(n\right)& {\underset{\_}{R}}_{3,3}\left(n\right)& {\underset{\_}{R}}_{3,4}\left(n\right)& {\underset{\_}{R}}_{3,5}\left(n\right)& {\underset{\_}{R}}_{3,6}\left(n\right)& {\underset{\_}{R}}_{3,7}\left(n\right)& {\underset{\_}{R}}_{3,8}\left(n\right)\\ {\underset{\_}{R}}_{4,1}\left(n\right)& {\underset{\_}{R}}_{4,2}\left(n\right)& {\underset{\_}{R}}_{4,3}\left(n\right)& {\underset{\_}{R}}_{4,4}\left(n\right)& {\underset{\_}{R}}_{4,5}\left(n\right)& {\underset{\_}{R}}_{4,6}\left(n\right)& {\underset{\_}{R}}_{4,7}\left(n\right)& {\underset{\_}{R}}_{4,8}\left(n\right)\\ {\underset{\_}{R}}_{5,1}\left(n\right)& {\underset{\_}{R}}_{5,2}\left(n\right)& {\underset{\_}{R}}_{5,3}\left(n\right)& {\underset{\_}{R}}_{5,4}\left(n\right)& {\underset{\_}{R}}_{5,5}\left(n\right)& {\underset{\_}{R}}_{5,6}\left(n\right)& {\underset{\_}{R}}_{5,7}\left(n\right)& {\underset{\_}{R}}_{5,8}\left(n\right)\\ {\underset{\_}{R}}_{6,1}\left(n\right)& {\underset{\_}{R}}_{6,2}\left(n\right)& {\underset{\_}{R}}_{6,3}\left(n\right)& {\underset{\_}{R}}_{6,4}\left(n\right)& {\underset{\_}{R}}_{6,5}\left(n\right)& {\underset{\_}{R}}_{6,6}\left(n\right)& {\underset{\_}{R}}_{6,7}\left(n\right)& {\underset{\_}{R}}_{6,8}\left(n\right)\\ {\underset{\_}{R}}_{7,1}\left(n\right)& {\underset{\_}{R}}_{7,2}\left(n\right)& {\underset{\_}{R}}_{7,3}\left(n\right)& {\underset{\_}{R}}_{7,4}\left(n\right)& {\underset{\_}{R}}_{6,5}\left(n\right)& {\underset{\_}{R}}_{7,6}\left(n\right)& {\underset{\_}{R}}_{7,7}\left(n\right)& {\underset{\_}{R}}_{7,8}\left(n\right)\\ {\underset{\_}{R}}_{8,1}\left(n\right)& {\underset{\_}{R}}_{8,2}\left(n\right)& {\underset{\_}{R}}_{8,3}\left(n\right)& {\underset{\_}{R}}_{8,4}\left(n\right)& {\underset{\_}{R}}_{6,5}\left(n\right)& {\underset{\_}{R}}_{8,6}\left(n\right)& {\underset{\_}{R}}_{8,7}\left(n\right)& {\underset{\_}{R}}_{8,8}\left(n\right)\end{array}\right]$

It will be appreciated that each block R_i,j(n) for i=1, . . . , 6 and j=1, . . . , 6 is a 5×5 matrix and cross-polynomial terms were not included in the vector u(n) as discussed above. Accordingly, these Type 1 and Type 2 blocks may be computed as discussed above, and only the computation of the blocks in the last two row and last two columns will be described in detail. The last two columns and last two rows have different sizes than the first 36 blocks. For example, diagonal blocks R_7,7(n) and R_8,8(n) are 12×12 matrices, and off-diagonal blocks R_1,7(n) and R_1,8(n) are 5×12 matrices.

From the symmetry property, only the upper triangular blocks within the last two columns need to be considered. These 15 blocks can be grouped into 4 further types where the blocks in each type have the same structure as set forth in table 8 below (it will be recalled that block Types 1 and 2 were discussed above):

TABLE 10
Type	Blocks			Example
#	in this type	Dim	#	Expression
3	R1,7(n), R1,8(n)	5 × 12	2	R1,7(n) = x(n) · yH (n)
4	R2,7(n),..., R6,7(n) and	5 × 12	10	R2,7(n) = x(n − 1) · yH (n)
	R2,8(n),..., R6,8(n)
5	R7,7(n), R8,8(n)	12 × 12	2	R7,7(n) = y(n) · yH (n)
6	R7,8(n)	12 × 12	1	R7,8(n) = y(n) · zH (n)

Type 3 Blocks

Consider block R_1,7(n) shown below:

$\begin{array}{c}{\underset{\_}{R}}_{1,7}=\underset{\_}{x}\left(n\right)·{\underset{\_}{y}}^H\left(n\right)=\left[\begin{array}{c}1\\ a\\ a^2\\ a^3\\ a^4\end{array}\right]\left[b_1^2b_2^2b_3^2\dots b_{12}^2\right]x\left(n\right)·x^{*}\left(n\right)\\ =\left[\begin{array}{ccccc}{b}_1^2& b_2^2& b_3^2& \dots & b_{12}^2\\ {\mathrm{ab}}_1^2& {\mathrm{ab}}_2^2& {\mathrm{ab}}_3^2& \dots & {\mathrm{ab}}_{12}^2\\ a^2b_1^2& a^2b_2^2& a^2b_3^2& \dots & a^2b_{12}^2\\ a^3b_1^2& a^3b_2^2& a^3b_3^2& ⋮& a^3b_{12}^2\\ a^4b_1^2& a^4b_2^2& a^4b_3^2& \dots & a^4b_{12}^2\end{array}\right]a^2\end{array}$

Hence all 60 elements are real-valued.

The following is the procedure performed at the weight estimator 12 according to an embodiment of the present invention to compute the elements in block R_1,7(n):

• 1. Compute the first row, using one scalar-scalar multiplication for each b_m²·a²
• 2. Compute the remaining rows sequentially, by multiplying a previous row by a using one scalar-scalar multiplication per element.

The computation complexity is summarized below in table 9:

	TABLE 11
	Operations	#Mult	#Add
	60 scalar-scalar multiplications	60	0
	Subtotal	60	0
	Total 2 blocks	120	0
	Baseline: 2 × 60 complex-complex	480	240
	multiplications

Type 4 Blocks

Consider block R_2,7(n) where a₁=|x(n−1)| shown below:

$\begin{array}{c}{\underset{\_}{R}}_{2,7}=\underset{\_}{x}\left(n-1\right)·{\underset{\_}{y}}^H\left(n\right)=\left[\begin{array}{c}1\\ a_1\\ a_1^2\\ a_1^3\\ a_1^4\end{array}\right]\left[b_1^2b_2^2b_3^2\dots b_{12}^2\right]x\left(n-1\right)·x^{*}\left(n\right)\\ =\left[\begin{array}{ccccc}{b}_1^2& b_2^2& b_3^2& \dots & b_{12}^2\\ a_1b_1^2& a_1b_2^2& a_1b_3^2& \dots & a_1b_{12}^2\\ a_1^2b_1^2& a_1^2b_2^2& a_1^2b_3^2& \dots & a_1^2b_{12}^2\\ a_1^3b_1^2& a_1^3b_2^2& a_1^3b_3^2& ⋮& a_1^3b_{12}^2\\ a_1^4b_1^2& a_1^4b_2^2& a_1^4b_3^2& \dots & a_1^4b_{12}^2\end{array}\right]x\left(n-1\right)·x^{*}\left(n\right)\end{array}$

Hence, only one complex-complex multiplication is computed for all 60 elements. The following is the procedure performed by the weight estimator 12 to compute the elements in block R_2,7(n):

• 1. Compute x(n−1)·x*(n), using one complex-complex multiplication. Note that the conjugate of this term may be computed and stored while computing R_1,2(n).
• 2. Compute the first row, by multiplying x(n−1)·x*(n) by b_m², using one scalar-complex multiplication per element.
• 3. Compute the remaining rows sequentially, by multiplying a previous row by a₁ using one scalar-complex multiplication per element.

The computation complexity is summarized below in table 10:

	TABLE 12
	Operations	#Mult	#Add
	1 complex-complex multiplication	4	2
	60 scalar-complex multiplications	120	0
	Subtotal	124	2
	Total 10 blocks	1240	20
	Baseline: 10 × 60 complex-complex	2400	1200
	multiplications

Type 5 Blocks

Consider block R_7,7(n) shown below:

$\begin{array}{c}{\underset{\_}{R}}_{7,7}=\underset{\_}{y}\left(n\right)·{\underset{\_}{y}}^H\left(n\right)=\left[\begin{array}{c}{b}_1^2\\ b_2^2\\ b_3^2\\ ⋮\\ b_{12}^2\end{array}\right]\left[b_1^2b_2^2b_3^2\dots b_{12}^2\right]x\left(n\right)·x^{*}\left(n\right)\\ =\left[\begin{array}{ccccc}{b}_1^2b_1^2& b_1^2b_2^2& b_1^2b_3^2& \dots & b_1^2b_{12}^2\\ b_2^2b_1^2& b_2^2b_2^2& b_2^2b_3^2& \dots & b_2^2b_{12}^2\\ b_3^2b_1^2& b_3^2b_2^2& b_3^2b_3^2& \dots & b_3^2b_{12}^2\\ ⋮& ⋮& ⋮& ⋮& ⋮\\ b_{12}^2b_1^2& b_{12}^2b_2^2& b_{12}^2b_3^2& \dots & b_{12}^2b_{12}^2\end{array}\right]a^2\end{array}$

All elements in the block are real valued. From the symmetry property, only the 78(=12·13/2) upper triangle elements need to be computed. For each of these elements, the term b_i²·b_j²·a² is computed, which requires 2 scalar-scalar multiplications. The computation complexity is summarized below in table 11:

	TABLE 13
	Operations	#Mult	#Add
	78 double scalar multiplications	156	0
	Subtotal	156	0
	Total 2 blocks	312	0
	Baseline: 2 × 78 complex-complex	624	312
	multiplications

Type 6 Block

Consider block R_7,8 shown below:

$\begin{array}{c}{\underset{\_}{R}}_{7,8}=\underset{\_}{y}\left(n\right)·{\underset{\_}{z}}^H\left(n\right)=\left[\begin{array}{c}{b}_1^2\\ b_2^2\\ b_3^2\\ ⋮\\ b_{12}^2\end{array}\right]\left[b_1^4b_2^4b_3^4\dots b_{12}^4\right]x\left(n\right)·x^{*}\left(n\right)\\ =\left[\begin{array}{ccccc}{b}_1^2b_1^4& b_1^2b_2^4& b_1^2b_3^4& \dots & b_1^2b_{12}^4\\ b_2^2b_1^4& b_2^2b_2^4& b_2^2b_3^4& \dots & b_2^2b_{12}^4\\ b_3^2b_1^4& b_3^2b_2^4& b_3^2b_3^4& \dots & b_3^2b_{12}^4\\ ⋮& ⋮& ⋮& ⋮& ⋮\\ b_{12}^2b_1^4& b_{12}^2b_2^4& b_{12}^2b_3^4& \dots & b_{12}^2b_{12}^4\end{array}\right]a^2\end{array}$

All 144 elements in the block are real valued. For each of these elements, the term b_i²·b_j⁴·a² is computed, which requires 2 scalar-scalar multiplications. The computation complexity is summarized below in table 12:

	TABLE 14
	Operations	#Mult	#Add
	144 double scalar-scalar	288	0
	multiplications
	Total	288	0
	Baseline: 144 complex-complex	576	288
	multiplications

Computation Complexity

The section will summarize the reduction in computation complexity, in the correlation matrix, for the improved methodology with cross terms. There are 1485(=54*55/2) elements in the upper triangular correlation matrix. The complexity is summarized below in table 13:

	TABLE 15
	Operations	#Mult	#Add
	Type 1 (diagonal) blocks (6)	60	6
	Type 2 (off-diagonal) blocks (15)	780	30
	Type 3 blocks (2)	120	0
	Type 4 blocks (10)	1240	20
	Type 5 blocks (2)	312	0
	Type 6 block (1)	288	0
	Total	2800	56
	Baseline: 1485 complex-complex	5940	2970
	multiplications

Cross Correlation Vector

Next, reducing the computation complexity for determining the cross correlation vector will be described. The cross correlation vector at time n is:

$\begin{array}{c}\underset{\_}{v}\left(n\right)=d\left(n\right)·\underset{\_}{u}\left(n\right)\\ =\left[\begin{array}{c}d\left(n\right)·\underset{\_}{x}\left(n\right)\\ d\left(n\right)·\underset{\_}{x}\left(n-1\right)\\ d\left(n\right)·\underset{\_}{x}\left(n-2\right)\\ d\left(n\right)·\underset{\_}{x}\left(n-3\right)\\ d\left(n\right)·\underset{\_}{x}\left(n-4\right)\\ d\left(n\right)·\underset{\_}{x}\left(n-5\right)\\ d\left(n\right)·\underset{\_}{y}\left(n\right)\\ d\left(n\right)·\underset{\_}{z}\left(n\right)\end{array}\right].\end{array}$ $\mathrm{where}$ $d\left(n\right)·\underset{\_}{y}\left(n\right)=\left[\begin{array}{c}{b}_1^2·d\left(n\right)·x\left(n\right)\\ b_2^2·d\left(n\right)·x\left(n\right)\\ b_3^2·d\left(n\right)·x\left(n\right)\\ ⋮\\ b_{12}^2·d\left(n\right)·x\left(n\right)\end{array}\right]=\left[\begin{array}{c}{b}_1^2\\ b_2^2\\ b_3^2\\ ⋮\\ b_{12}^2\end{array}\right]d\left(n\right)·x\left(n\right)\mathrm{and}$ $d\left(n\right)·\underset{\_}{z}\left(n\right)=\left[\begin{array}{c}{b}_1^2·b_1^2·d\left(n\right)·x\left(n\right)\\ b_2^2·b_2^2·d\left(n\right)·x\left(n\right)\\ b_3^2·b_3^2·d\left(n\right)·x\left(n\right)\\ ⋮\\ b_{12}^2·b_{12}^2·d\left(n\right)·x\left(n\right)\end{array}\right].$

For 0≦l≦5, the computation of each cross correlation vector block d(n)·x(n−l) was discussed in detail above. Accordingly, on the computation of the cross correlation vector blocks with cross term will be described in detail.

First d(n)·y(n) is computed, and then is used to compute d(n)·z(n). To compute d(n)·y(n), a complex-complex multiplication is used once to compute d(n)·x(n). Then, the elements b₁²·d(n)·x(n) and b₁²·b₁²·d(n)·x(n) for 1≦l≦12 are obtained using scalar-complex multiplication.

The following is the sequence performed by the weight estimator 12 according to an embodiment of the present invention to compute the cross-correlation vector for y(n) and z(n) sections:

• 1. Compute d(n)·x(n), using complex-complex multiplication.
• 2. Compute b₁²·d(n)·x(n) and then b₁²·b₁²·d(n)·x(n), using scalar-complex multiplication, for 1≦l≦12.

The computation complexity is summarized below in table 14:

	TABLE 16
	Operations	#Mult	#Add
	x(n) sections
	6 complex-complex	24	12
	multiplications
	24 scalar-complex	48	0
	multiplications
	y(n) and z(n) sections
	1 complex-complex	4	2
	multiplication
	24 scalar-complex	48	0
	multiplications
	Total	124	14
	Baseline: 54 complex-complex	216	108
	multiplications

The present invention provides methodologies to reduce the computation complexity for digital pre-distortion. As demonstrated above, more than half the correlation computation may be reduced.

The invention being thus described, it will be obvious that the same may be varied in many ways. For example, while mentioned as applicable to digital pre-distortion for RF amplifiers, the present invention is also application to pre-distortion for other types of amplifiers. Furthermore, the computation reduction techniques discussed above for the different types of blocks and cross correlation vectors may be used together or may be used independently of each other. In other words, the computation reduction technique for only one type of block may be used. Such variations are not to be regarded as a departure from the invention, and all such modifications are intended to be included within the scope of the invention.

Claims

1. A method of pre-distorting a signal for input to an amplifier, comprising: generating a weight vector based on a pre-distorted input signal and an output signal of the amplifier using a computation reduction technique; andpre-distorting an input signal for input to an amplifier based on the generated weight vector; andwherein the computation reduction technique decomposes a number of multiplication operations between complex numbers such that a total number of multiplications and a total number of additions to generate the weight vector is reduced as compared to if the number of multiplication operations between complex numbers was not decomposed.

2. The method of claim 1, wherein the generating step comprises: computing a correlation diagonal block derived from one of the output signal and a previous output signal, a first term of the vector being a complex number and having a magnitude, and each subsequent term of the vector being obtained by multiplying the first term by a result of raising the magnitude to a respective exponential order.

3. The method of claim 2, wherein the computing step comprises: computing a square of the first term to obtain a first scalar value;computing a square root of the square to obtain a second scalar value;using scalar multiplication and the first and second scalar values to obtain elements in the diagonal block.

4. The method of claim 1, wherein the generating step comprises: computing a correlation off-diagonal block derived from a first vector and a second vector, the first vector derived from one of the output signal and a previous output signal, a first term of the vector being a complex number and having a magnitude, and each subsequent term of the first vector being obtained by multiplying the first term by a result of raising the magnitude to a respective exponential order, andthe second vector derived from a previous output signal, and a first term of the second vector being a previous first term of the first vector.

5. The method of claim 4, wherein the computing step comprises: computing a first element of a first row of the off-diagonal block as the first term of the first vector times a complex conjugate of the first term of the second vector;computing remaining elements of the first row of the off-diagonal block using scalar-complex multiplication, the first element and a magnitude of the first term of the second vector;computing each subsequent row of the off-diagonal block using scalar-complex multiplication, a previous row of the off-diagonal block, and the magnitude of the first term of the first vector.

6. The method of claim 1, wherein the generating step comprises: computing a cross correlation vector block derived from the pre-distorted input signal and a vector, the vector obtained from one of the output signal and a previous output signal, a first term of the vector being a complex number and having a magnitude, and each subsequent term of the vector being obtained by multiplying the first term by a result of raising the magnitude to a respective exponential order.

7. The method of claim 6, wherein the computing step comprises: computing a first element of the cross-correlation vector block as the pre-distortion input signal times the first term; andcomputing a number of subsequent elements of the cross-correlation block using scalar multiplication a previous term and the magnitude of the first term.

8. The method of claim 1, wherein the generating step comprises: computing an off-diagonal cross correlation block derived from a first vector and a second vector, the first vector derived from one of the output signal and a previous output signal as a first term, a first term of the vector being a complex number and having a magnitude, and each subsequent term of the vector being obtained by multiplying the first term by a result of raising the magnitude to a respective exponential order, andeach term of the second vector derived from a different previous output signal and the output signal.

9. The method of claim 8, wherein the computing step comprises: computing a square of the first term to obtain a first scalar value;computing a square of each term in the second vector to obtain a set of second scalar values; andusing scalar multiplication, the first scalar value and the set of second scalar values to obtain elements in the off-diagonal cross correlation block.

10. The method of claim 1, wherein the generating step comprises: computing an off-diagonal cross correlation block derived from a first and a second vector, each term of the first vector being a different previous output signal squared times the output signal, andeach term of the second vector being a square of a corresponding term in the first vector.

11. The method of claim 10, wherein the computing step comprises: computing a square of each different previous output signal to obtain a first set of scalar values;computing a square of each value in the first set of scalar values to obtain a second set of scalar values;computing a square of the output signal to obtain a third scalar value; andusing scalar multiplication, the first set of scalar value, the second set of scalar value and the third scalar value to obtain the elements of the off-diagonal cross-correlation block.

12. The method of claim 1, wherein the generating step comprises: computing a diagonal cross-correlation block derived from a first vector, each term of the first vector being a different previous output signal squared times the output signal.

13. The method of claim 12, wherein the computing step comprises: computing a square of each different previous output signal to obtain a first set of scalar values;computing a square of the output signal to obtain a second scalar value; andusing scalar multiplication, the first set of scalar values and the second scalar value to obtain the elements of the diagonal cross-correlation block.

14. The method of claim 1, wherein the generating step comprises: computing a cross correlation vector block derived from the pre-distorted input signal and a first vector, each term of the first vector being a different previous output signal squared times the output signal.

15. The method of claim 14, wherein the computing step comprises: compute the pre-distortion input signal times the output signal to obtain a complex value;compute the square of each different previous output signal to obtain a set of scalar values; andusing scalar-complex multiplication, the complex value and the set of scalar values to obtain elements of the cross correlation vector block.

16. The method of claim 1, wherein the generating step comprises: computing a cross correlation vector block derived from the pre-distorted input signal and a first vector, each term of the first vector being a different previous output signal quadrupled times the output signal.

17. The method of claim 14, wherein the computing step comprises: compute the pre-distortion input signal times the output signal to obtain a complex value;compute the square of each different previous output signal to obtain a set of scalar values; andusing scalar-complex multiplication, the complex value and the set of scalar values to obtain elements of the cross correlation vector block.