An analysis of real-Fourier domain-based adaptive algorithms implemented with the Hartley transform using cosine-sine symmetries
Summary (2 min read)
Introduction
- THE ability of an adaptive filter to operate effectively inan unknown environment and track time variations of the input statistics makes it a powerful device.
- The authors exploit the symmetry of the Hartley transform in being decomposed into a cosine and a sine transform of reduced order and the symmetrical structure of the HBNLMS to significantly reduce the computational complexity of the FBNLMS.
- The authors also note that the HBNLMS requires almost the same order of memory as that of the FBNLMS implementation.
A. Notation
- The authors assume real inputs, real “desired” outputs and they start with a real .
- This will ensure that the weight vectors over all iterations are real.
- The operation refers to elementwise multiplication of two sequences.
- The superscripts in and refer, respectively, to “circulant” and “circulant symmetric.” , , and refer to the Fourier, the Hartley, the inverse Fourier and the inverse Hartley matrices, respectively.
- The filter output y(n) is given by the convolution sum (1) This output y(n) is used to estimate the desired response d(n).
II. FBNLMS ALGORITHM
- The Fourier transform-based block normalized least mean square algorithm [12], [13] is an efficient way of im- plementing the block normalized least mean squares algorithm [12].
- The following algebraic manipulations then follow trivially: (8) It can be seen that diag , where the authors exploit the realness of the input vector and the fact that Fourier matrices diagonalize circulant matrices [14].
- This, therefore, leads to (9) where the symbol denotes the elementwise multiplication of the two N-point sequences.
- Equation (11) can then be rewritten similarly as (12) Authorized licensed use limited to: Eindhoven University of Technology.
- The authors would like to comment that, since the FBNLMS algorithm is an exact transformation of the BNLMS algorithm, the dynamic behavior of both of them will be essentially the same [12].
III. HARTLEY TRANSFORM-BASED BNLMS (HBNLMS)
- The HBNLMS would be computationally disadvantageous when compared to the FBNLMS (if the symmetries in the sequences are not used), as is done in [8], where a Hartley transform of order 2N is implicitly used.
- The symmetry in the sequences helps us in replacing a 2N-point Hartley transform with cosine transforms of order of N/2 down until 1 [17], [18].
- The DST-II of order L can be computed from the DCT-II of the alternate sequence as shown below: (26) Authorized licensed use limited to: Eindhoven University of Technology.
V. DCT AND DFT COMPARISONS
- The algorithm to be used for the DFT depends on the application Authorized licensed use limited to: Eindhoven University of Technology.
- Instead of going into the specifics of algorithms and applications, the authors use the radix-2 versions of DCT and DFT to illustrate the advantages of the two implementation schemes.
- Their choice for the real algorithms is restricted to the radix-2 counterparts on both sides since their motive is in showing the relative advantages rather than comparing the individual algorithms.
- To find the DCT-I of and the DCT-I of the 2N-point sequence, an additional (3/2 N) additions are required.
- The savings in Table II are computed with respect to multiplications, as they are more prominent for small N. For , using the FBNLMS, the authors require 3 727 390 additions and 1 433 620 multiplications, for a total of 5 161 010 computations.
VI. MEMORY REQUIREMENTS
- Computation count is just one of the many parameters that reflect the ‘desiredness’ of an algorithm.
- The algorithm would not serve much purpose if the memory requirements of the same were to far exceed that of the original algorithm.
- The memory requirements of the FBNLMS are the incoming data, (length N) and (length M), two buffer registers of length N to store the intermediate results (like forward transforms etc.), a desired data register of length B, and twiddle factors that can be stored in a register of length less than N if the authors use their complex conjugate symmetry.
- Unit register locations are needed to store the adaptation constants and the variance estimator.
- Thus, the HBNLMS requires an order of memory units.
VII. CONCLUSION
- The HBNLMS has been implemented using the DCT-DST symmetric decomposition.
- Many computation reduction techniques that can be employed with the DFT, like higher-radix and split-radix algorithms, pruning [24]–[26], block transforms [27], and slide transforms [28] can be easily extended to the DHT, DCT and DST, thereby resulting in an almost equivalent reduction in complexity of the HBNLMS over the FBNLMS with such techniques.
- Besides, the memory requirements of the HBNLMS are of the same order as that of the FBNLMS.
- The stability of the HBNLMS algorithm vis-à-vis rounding off errors can Authorized licensed use limited to: Eindhoven University of Technology.
Did you find this useful? Give us your feedback
Citations
211 citations
83 citations
33 citations
Cites result from "An analysis of real-Fourier domain-..."
...Moreover, the proposed may be employed in adaptive algorithms in a way similar to that reported in [4]....
[...]
15 citations
9 citations
References
16,062 citations
724 citations
661 citations
"An analysis of real-Fourier domain-..." refers methods in this paper
...We do not use an explicit dependence of the weight vector and the error vector on the discrete-time variable “n” as we are dealing with a particular iteration, and for that iteration, these vectors can be assumed to be constant....
[...]
489 citations
"An analysis of real-Fourier domain-..." refers methods in this paper
...The filter output y(n) is given by the convolution sum (1) This output y(n) is used to estimate the desired response d(n)....
[...]
...In the block algorithm (where B is the filter block length), we define the block input (an M B matrix) as (2) Then, the block output is defined as , where (3) The estimation error is defined as (4) refers to the adaptation constant in the adaptive algorithm....
[...]
483 citations
"An analysis of real-Fourier domain-..." refers background in this paper
...This, therefore, leads to (9) where the symbol denotes the elementwise multiplication of the two N-point sequences....
[...]
Related Papers (5)
Frequently Asked Questions (14)
Q2. What is the main motive of this work?
The main motive of this work is to exploit the symmetries inherent in the sequences to use the computationally more efficient DCT and DST systems to recursively compute the needed transforms.
Q3. What are the memory requirements of the FBNLMS?
The memory requirements of the FBNLMS are the incoming data, (length N) and (length M), two buffer registers of length N to store the intermediate results (like forward transforms etc.), a desired data register of length B, and twiddle factors that can be stored in a register of length less than N if the authors use their complex conjugate symmetry.
Q4. What is the function of the block normalized least mean square?
The Fourier transform-based block normalized least mean square (FBNLMS) algorithm [12], [13] is an efficient way of im-plementing the block normalized least mean squares (BNLMS) algorithm [12].
Q5. What is the computation count of an algorithm?
The computation count of an algorithm is defined as the number of multiplications and additions, since in most modern-day digital signal processors (DSPs), the time needed for implementing an addition is almost the same as that for a multiplication.
Q6. What is the disadvantage of the HBNLMS?
The HBNLMS would be computationally disadvantageous when compared to the FBNLMS (if the symmetries in the sequences are not used), as is done in [8], where a Hartley transform of order 2N is implicitly used.
Q7. What is the value of the adaptive algorithm?
In the block algorithm (where B is the filter block length), the authors define the block input (an M B matrix) as(2)Then, the block output is defined as , where(3)The estimation error is defined as(4)refers to the adaptation constant in the adaptive algorithm.
Q8. What is the basic idea of the HBNLMS?
The basic idea of their implementation is that the recursive implementation of the DCT-I, DST-I, DCT-II and DST-II involve the same building blocks with signs changed appropriately to give the needed transforms.
Q9. What is the inverse Hartley transform of order 2N?
Since the inverse DST-II can be computed using the inverse DCT-II [9], the inverse Hartley transform of order 2N actually needs the same order of computations as a forward transform of length N. can be computed in exactly a similar way as , as is similar in structure to .
Q10. How can the authors use the FBNLMS algorithm?
The FBNLMS algorithm can then be implemented by extending all data matrices to be circulant and symmetric and using to diagonalize them [8].
Q11. What are the advantages of the HBNLMS?
Many computation reduction techniques that can be employed with the DFT, like higher-radix and split-radix algorithms, pruning [24]–[26], block transforms [27], and slide transforms [28] can be easily extended to the DHT, DCT and DST, thereby resulting in an almost equivalent reduction in complexity of the HBNLMS over the FBNLMS with such techniques.
Q12. What is the DST-II of an alternate sequence?
the 2N-point Hartley transform of is(27)Since the DCT-II of x(n) is implemented recursively using an even and an odd breakup [21], [22], the DCT-II of (and thus the DST-II and the DST-I) can be easily computed.
Q13. What is the HBNLMS's requirement for memory units?
The HBNLMS, on the other hand, has to incorporate the inherent symmetry of the data sequences if the authors want to reduce the number of memory units used.
Q14. How many computations are required for the FBNLMS?
The savings in Table II are computed with respect to multiplications, as they are more prominent for small N.For , using the FBNLMS, the authors require 3 727 390 additions and 1 433 620 multiplications, for a total of 5 161 010 computations.