>Emre,
>Thanks for the reply. Actually it should be possible to attach files
>with NNTP but I think my ISP is preventing attachments. Anyway I put
>the image at http://www.dspia.com/receiver/eqout.jpg .
>As you can see the equalizer output is not centered around 1,1
>(although it's centered around zero) and not normally distributed
>around 1,-1 (although it mostly is around zero).
>The input to the equalizer is coming through a linear frequency
>selective channel so I am not sure if the PSD of that signal is
>meaningful. The input to the channel is a 3 level signal where the
>probability of 1, -1 are 1/4 each and the probability of 0 is 1/2.
>I have tried increasing the update error by 2 for 1,-1 sliced cases to
>accomodate the lower probability of symbols but it only helped very
>little. Any suggestions on how to improve this would be appreciated.
I am not sure what is wrong with the LMS filter. It might be that
minimizing the mean squared error introduces a bias towards 0. At any
rate, if you can't fix the filter but can estimate statistics for each
symbol {-1,0,1} separately, then you could adjust your slicing points to
account for the skewed distribution.
Hope this helps,
Emre
Reply Start a New Thread
kal@dspia.com wrote:
...
> Thanks for the reply. Actually it should be possible to attach files
> with NNTP but I think my ISP is preventing attachments.
Images are forbidden on usenet except in binary newsgroups. Some ISPs
ignore this rule. Others enforce it in various ways. Because some ISPs
take no measures to prevent binary content, others remove the image or
simply suppress the entire message.
Jerry
--
Engineering is the art of making what you want from things you can get.
�����������������������������������������������������������������������
Reply Start a New Thread
On Wed, 19 Nov 2008 09:28:18 -0600, "emre" <eguven@ece.neu.edu> wrote:
>Two things: (i) Your image is not visible, nor is there a way to post an
>image to a post here. Maybe you can post a link to it. (ii) The
>distribution of your quantized values depends on the pdf of the input. If
>you had an input that is uniformly distributed between [-1.5, 1,5], then
>you would get equally likely quantized values of {-1,0,1} using -0.5 and
>0.5 as slicing points.
>
>Your question relates to the problem of quantization. If you know the
>input pdf and you can define a criterion for the output, such as obtaining
>a uniform distribution or minimizing MSE, then you can get adjust the
>quantized values and/or slicing points to meet this objective.
Emre,
Thanks for the reply. Actually it should be possible to attach files
with NNTP but I think my ISP is preventing attachments. Anyway I put
the image at http://www.dspia.com/receiver/eqout.jpg .
As you can see the equalizer output is not centered around 1,1
(although it's centered around zero) and not normally distributed
around 1,-1 (although it mostly is around zero).
The input to the equalizer is coming through a linear frequency
selective channel so I am not sure if the PSD of that signal is
meaningful. The input to the channel is a 3 level signal where the
probability of 1, -1 are 1/4 each and the probability of 0 is 1/2.
I have tried increasing the update error by 2 for 1,-1 sliced cases to
accomodate the lower probability of symbols but it only helped very
little. Any suggestions on how to improve this would be appreciated.
Reply Start a New Thread
>Hi everyone,
>I'm working on an LMS adaptive filter which is behaving in somewhat of
>an unexpected way. The attached image shows the histogram of the
>equalizer output after convergence; my slicer thresholds are at 0.5
>and -0.5 (ie anything above 0.5 is sliced to 1, anything below -0.5 is
>sliced to -1 and anything in between is sliced to 0; this is a 3 level
>code.) Then I use an error generated from this sliced value and the
>actual output to adapt the filter.
>As one can see the mean of the outputs close to 1 and -1 are not
>really 1 and -1 which is what I expected. Also the distributions of
>the outputs around 1,-1 are not really symmetric; again something I'd
>expect to see. Obviously what I am trying to do is to make the
>histogram more coalesced around 1,0,-1 to get a more distance between
>decision points.
>I know the question is a little vague without a lot more detail but
>does anyone have any comments on why this might be happening?
>
>Thanks.
>
>Muzaffer Kal
Muzaffer,
Two things: (i) Your image is not visible, nor is there a way to post an
image to a post here. Maybe you can post a link to it. (ii) The
distribution of your quantized values depends on the pdf of the input. If
you had an input that is uniformly distributed between [-1.5, 1,5], then
you would get equally likely quantized values of {-1,0,1} using -0.5 and
0.5 as slicing points.
Your question relates to the problem of quantization. If you know the
input pdf and you can define a criterion for the output, such as obtaining
a uniform distribution or minimizing MSE, then you can get adjust the
quantized values and/or slicing points to meet this objective.
Hope this helps,
Emre
Reply Start a New Thread
On Tue, 18 Nov 2008 23:27:42 -0800, Muzaffer Kal <kal@dspia.com>
wrote:
Here is the image.
����
$.' ",#(7),01444'9=82<.342��
2!!22222222222222222222222222222222222222222222222222��
��
%&'()*456789:CDEFGHIJSTUVWXYZcdefghijstuvwxyz���������������������������������������������������������������������������
��
$4�%�&'()*56789:CDEFGHIJSTUVWXYZcdefghijstuvwxyz��������������������������������������������������������������������������
(��
(��
(�څ�Zf�5��q.v�71�g�8���+?0��K%��rnar���0.�r1�p�=#KS�t}����ll<��k�H����3����?04(��3]���7�'U���q�}��%ٜ�;I�pz�Р�(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��(��+�Ş�/$����iu7�=�q��2��d
=źA*O,-r�)#&C\�nz��^��zEՈ1)�2�ʌ�3܅eo�?�"���A�����f���ng���dV�<Ѥi�w
.���ώ�K}SO��?t��X�YT7�h23�F}�u����Cѿ��
�m��A���yHp>_RO�]]R��
��
��O�
��
��O�
(��
(��
(��
(��
(��
(��
(��
(��
�c�
*�YS����#_�DQEjlQE
(��
(��
(��
(��
(��
(��
(��
��Ǥ���?�?z�T�������g�
(��
(��
(��
(��
(��
(��
(��)k���
�״?��v�E�
(��
(��
(��
(��
d�G</Ѥ�H�]eX ��(>G���(t��r����z�
(��
(��
(��
(��
(��
(��
(��)_�ޛ�
*�T��
Q�n�����
�״?��v�������h|/��l(����(��
Z��?/�
*�YS����#_�DQEjlQE
ʡ��A?�3�h�N����4-D��t�;g�-�-��3��'���
�״?��v�E�
��
(��
(��
(��
(��
(��
(��
(��
���
Z�YQ�~o�f4>��
��
#B����-�$��\�{^����Z$��eޠ�L�S�D��i�1�c�u�#���_h�-?��cjő��Vf$�>c̃ >�@����=��W�
/ڜ���-�*�3�v�u�_�?�����3S�*�Q=\���?Fd�6qŤAr����~DK�O�v%�v>��9���F�
s"O�H�L�2��Ա�u���t��Ӽnnm����R!B�e�VǯM�#��Qc�����
*�YS����#_�DQEjlQE
(��
(��
(��
(��
(��
(��
(��
(��
�c�
ޟ�:�
*�YS����#_�DQEjlQE
�y�
(��
(��
(��
(��
(��
(��
(��
��Ǥ���?�?z�T�������g�
(��
(��
(��
(��
(��
(��2u� �����M=r^�
(��6
(��
(��
(��
(��
(��
(��
(��)E�
��=�
(��
(��
(��
(��
��F��WQ5��J�2���?��.����tP��¾?�I���̾}�[+\OKb�ht26
W�Z$]�.�|����n��m-)#���OA]es���P�o�_�
3F��<߳K}'�?�k����1�m���=hQE
(��
(��
(��
(��
(��)_�ޛ�
*�T��
Q�n�����
�״?��v�������h|/��l(����(��
Z�
��
���#�K�apJ�*����?z���?��n/a���P���+���S���*��*z
�?Bq����E���K��Py���/�\�`Õ ��Cu��*�P����}��f�G�n��J`u'����龿��2�V���L��>"�V��?��h�_ih|��LwrI�^�o�}�/��3�a����'��K�KB�-'�ՓF*lv>IP
(��
(��
(��
(��
(��
(��
(��
(��
�c�
�F�m�
�
��
7�.�'�
�[?�w��=��DJO�g�g�ҵ�]�������=��e�mc</?����Q?�?D�x[���_��V������^�?�뇴��9�qӚ�+��}�
(��
(��
(��
(��
(��)_�ޛ�
*�T��
Q�n�����
�״?��v�������h|/��l(����(��
��
��
(��
(��
(��
(��
(��
(��
(��
��Ǥ���?�?z�T�������g�
(��
(��
(��
(��
(��
�e���G�kF|���H�?�3��!����T�
�״?��v�E�
��
Ʋ��=,�~�Χ�Z^�m?���E$ӡ�6�X�etq�R&pA����|9�
�״?��v�E�
��
z�{�h�[�wz{?2����1m��|�q�s�<z���(��3N���~Ӭ�j[��i�x�q�F?s�=1�t(QE
3t2}HR+����2o��
#�y��̥���
�^��
��
�8��!Da?�sI�?�H�= �����������?�&��?9]�،m%t��J�
3�W�� �����M=b�,pxI�E�(����p���I=a'���=JMRv���
����e�Z��W� ԉP�iV�=�J��� ~=�����������Dzi4�a?7gw�zŎx����N3�Vw�k��!����qwg~bI3����Vt�sI����n?�P�Q�N��J ~7�?��7k�
*�WO◯�?��?_�QZ�Q@Q@Q@Q@Q@Q@Q@��
�?c�Y����s>X�7\��5�^|�?��:��?��5��]�~�I�9��?8�I|�C+f�p��&�=�~�G_��+,��?�osն�x���
���(nOu���]"8�����|�U02`������?|�wP�d�o�Iq?�=�wWrI�.Z[K�O����Ҧ����������ӥJJҋw�>�'��7��
(��69/�/��W�����uF��fS������c��y<1Ȳ[^^߷��B=��T��h��'�?�
�� '����?�$:7���
���zv�u}2�Em��X��g��sy��?�+�R��>R�eQ�;m�=1��$
���5����
���k�F
k5�$|�!�?��z�w�^�=�6Vv�����w�K%�����>����<s��V����[�f��,�JP_yW[�7?�&��%�
�״?��v�E�
��
�瘾�巕y�I�8��i�c�m%���,���on@�!�,XYX��P�t8�iqG?.��5�(�uDA�P!�
����!�
�?p���
�E��-%���
(��
(��
(��
(��
���
�`����3}kF��yc:ψa)�ocf@~`
� =��J�=W��ޔo �l���a�5���C�k3�X3B4)>�00yHNv���o�4_����
�s�?�*3�^]�i�m����;��|��\�eP}������t����.���?��h���;�v��=waFq���\B�5�������8�z&������?C�n����I�{���`��'tQǢ��h�1૩�"�c��%��wC
Ic32�ssk�8`�0q����y&���ڴ���5�2*�3�w6?..s?}��|G�
�Kmme2RWK������8��N-gk}V��K~�kl��ݽF@������l�ү<Ql���XK��
Q�ף���>��.҂t��/�����G�j�u�(�ɷ���$�}�y����3������]ia�bs(
l��,*W��F���ʱ�8�F�|��
Q�n�����
�״?��v�������h|/��l(����(��
��
��?����
��
?�
��@�
9�?���2�>��^|?>�=|���R��V_+��zG���YxsK��M��ir.A�ʀ��QZ4Q[�ec˩79����QL����
���k>!�F�V�?Y����? �?�}k*����gv
��U%n��_�s��㲼��m�McX���Y��[<>jL�S�olA�]�yΕ���|��`�N����X�k�|��?�������-��n���ê���~IQ[�XQE
��LG���sY??��ؙ�:��k��
��hA99?�Vy،C?Ӌ�
3�
�K 9�n��ǵX�?�"Qqn/tQE
(��
(��
(��
(��)E�
��=�
(��
(��
(��
(��
(��
��?����
��
(��
(��
(��
(��
(��
���H�I�b4�8'�v�?V`?���Me��R���O��F�k*Ed�B�G�!7�c�~��A��9߿�M�q�s�b�s�M�W:��kR�[���/�:�����9_k�Hc�`?���*��Ҹ-�B~�
�y��C��G,�88T��o����ㆩ/�I|ے_���ma��5+K����Qȹ#r�䀌�z����������
Ȍ�H
(��)k���
5�8<�v�
(���
(��
(��
(��
(��)E�
��=�
(��
(��
(��
(��
����$�|�8�6�ڸ���#-���#�c�}'��G�L�>�?dR툻�����?:�Վ227���)|u�.�wim?�e�g��iԯ�i�U�9�9�x玲����(z7����
(��
(��
(��
(��
(��
(��
(��
�c�
(��6
(��9�X�
Yóo]��9���
R�
��
)8��sr��?�<e�����%��~Y'��@�?�gV/��v�
�״?��v�E�
Y���}x�]Es��
&=2�V ��D�|��H�Y<�
� ߾���siW6�%�ZR�?�sy�*ޤ:�?}��Qvz��_��Q�V��n���z����^G1cy!�K�"���LbG����
m9{?r�i�G�����z�3���kJijqfu�)%M����}�O�4��|?����Ot���
(��
(��
�<Q�G����խ8cM�7)?M�nی^z����dRO4��I,����2�O?
��
X�K(h�y�im�E���ق\局22Mvz����CQ�?��?$�L�V�����A�}iB����e���{5�?����
(��
(��
(��
(�?�����d��Ks
�� <ii6s��|��$y��.�*�K��E����Z��om�V�g�d�+��=?~f�+���)�B*(1x��kJ����.QEY�QE
|��"��V��Hd���`�*����zp��k/��md�����3m.�Mp,�����X�$e
�Z[��2�J��?-/��
QЯ���y/��/�ZM��7gn��=��RM��)J4�!t����u�o]�4��t�X�l���8?m�8ł+*?�%�Cc�?^���g?������[B?�s�!@8��p�����Q�
���f���in���X�\?��'���h����NnS]5
�:Z��{[�{X��Z\G,������{��0�I�y֢��:������V<��:m�������?��j
�
?���<�m�W�F�?��[M]�}3��3�V{X]��\�9��?�?���4hW�o�K�
Q�n�����
�״?��v�������h|/��l(����(��
"�
(��
*��y%��qq�\��n�>W��X���Q�u�:�6��Ķ{9,�)-%?.d�Ѳ#f?�w-�g1?Ԁ�)�;?p�kY�H�f���ϟqX?�;�.q���22{dTW�?in..o�a��bI2����'Xc��W�j������`h-›)%(�k|̑L���vBH%��i�:/�N����C�n|�e"���-�^F?��a�9��@cowd?G�)B�=Ij�it[��Y�2ߋ<F�f�m{k���X���Ϋ!��p9얓��Wܫ/,��O���f����1��\�&�:3�L`�`U�B�����7�4��E#�"�r9]�vs���rKdk�����~����?쿽����F�3�x�zJ�n]M%��R?8Qw�j��g{��~�+q��'��
�s�?�*3�\���~������C�ۼ�F�(����?+�d�U�x�ÒO�i��j
����>Yd�<�wcv�8'
}2�?���
Z��?/�
�s��z�M��/�����
�4k��z��~.^���o�vٶ�#�F�0��u���ml��O�>)c��H�P��r� �:���_RO�]�?�?��?{O�͞FL����ki�������/G��;`F�ln]�m<i�E�2�6�B{�������|?7}�۽�Y���郞V_�?�?qz��):����Ԯc#
�-��
(��6
(��+����+0U�)?S�cuq�eF}�#Ʃt|!�Ke"�sl�t�� ?d�9?'
W�Մ��Z��
�-�������Ge��!�x �x��;/W���3�A;6�������=��f�����?�̥�z�?|TW����⯼E|�������N�\�*��B�9*FܒѮ:?(`0*�$���n?�M?�͜"U�>�ipzev0p�A�1V|/�|�fK�~�iw$�Q[#Ȇ6h6��N�0O�\�Fk�?l�S@�o߮���ޞo��&�[���x�1��'�7k��
�MGB��UYnm���(,��g�5z�t��?�V�
���TMzE|EQYYI����&�Of�5���V]b��RVfj?<���y�bh���8 ?�Z&��8�S�9�KtQE2
Q�n�����
�״?��v�������h|/��l(����(��
���=N��rNI$��zt����
����
(��
(��
(��
(��
(��
(��
(�Լ5��F�[?SD�Ō�K�[x�� �,
����,?�c&�C���ۊ�)� ��
Q�T��A#s]6������YfU�d
�*�=p@8�v�f�gi(�����^Y�{F(e;v?�@B�1�q'?�`�w�V��_j�d����w��ej����ܤ��<��
����
����
����
����
����
����
����
����
=s��ùO�O�
�-��$O�}��dr�;r:��zTZ����5���4Qjde�cej�0��l�a�� V֗�jZu�������M9ah�6�Y���<�n��`��G�g�[�o'����_6i-��(HP�
Y�p�$���
Cn*�{n�
�o�H�(8��wĔQE
�i?c/z�^H���&�mb��V����}���la�8%y# ��?x�4�K��a���i��v��Q��G�>o��|����5��x�4����X�&M.���±wHw�d��'��j?�����=��k��^귚e����%?�#��`?���?B}��
(��
(��
(��
(��
(��
(��
(��
(��
(��
(��
(��
(��
(��
(��
(��
(��
(��
(��
(��
(��
(��
(��
(��
(��
(��
(��
(��
(��
(��
(��
(��
(��
(��
(��3�֬ ա�%���n#�/�� ��cf�v�vq�Ь;�5�?ٲZ�ͥ@w��\8�>,&ӀX��{[�-?�QE
Muzaffer Kal
DSPIA INC.
ASIC/FPGA Design Services
http://www.dspia.com
Reply Start a New Thread
Hi everyone,
I'm working on an LMS adaptive filter which is behaving in somewhat of
an unexpected way. The attached image shows the histogram of the
equalizer output after convergence; my slicer thresholds are at 0.5
and -0.5 (ie anything above 0.5 is sliced to 1, anything below -0.5 is
sliced to -1 and anything in between is sliced to 0; this is a 3 level
code.) Then I use an error generated from this sliced value and the
actual output to adapt the filter.
As one can see the mean of the outputs close to 1 and -1 are not
really 1 and -1 which is what I expected. Also the distributions of
the outputs around 1,-1 are not really symmetric; again something I'd
expect to see. Obviously what I am trying to do is to make the
histogram more coalesced around 1,0,-1 to get a more distance between
decision points.
I know the question is a little vague without a lot more detail but
does anyone have any comments on why this might be happening?
Thanks.
Muzaffer Kal
DSPIA INC.
ASIC/FPGA Design Services
http://www.dspia.com
Reply Start a New Thread