In certain situations, the underlying source distribution
or the distortion function
of the source variable varies during the quantization process. It is practically infeasible to design separate codebooks optimized for every different source distribution and distortion function, or the encoder and the decoder may not have the ability to store a large number of codebooks. In these situations, it is convenient to use a quantizer whose codebook is constructed by a transformation of a fixed codebook based on the current statistical distribution of the source variable. These types of quantizers are generally called transformed quantizers [
22,
23], and have been used in the conventional source coding area with a linear orthogonal transformation followed by a product quantizer. We provide in this subsection an analysis of the generalized vector quantizer, which is described in Section 2, when a transformed codebook is used. Detailed applications to finite-rate feedback MISO systems with a transformed codebook over spatially correlated fading channels are provided in Section 5.
It is first assumed that all the codebooks are generated from one fixed codebook
, which is designed to match the source distribution
, and distortion function
with sensitivity matrix
. Codebook
has a point density given by
, and a normalized inertial profile
that is optimized to match the distortion function
, with
representing the asymptotic Voronoi cell that contains
with side information
. Let the source distribution change from
to
, and let the distortion function become
instead of
with sensitivity matrix
instead of
. The encoder and decoder are assumed to adapt a transformed codebook
obtained from
by using a general one-to-one mapping
with both its domain and codomain in space
, that is,
3.2. Suboptimal Point Density and suboptimal Voronoi Shape
Assuming the codebook transformation function
has a continuous first order derivative, two types of suboptimality arise when the transformed quantizer is used. One comes from the suboptimal point density
, which can be derived from
as
If the source variable is subject to
constraints given by vector equation
, the transformed point density is given by
where
is an orthonormal matrix whose columns constitute an orthonormal basis for the null space
. Compared to the optimal point density
given by (8), which corresponds to the optimally designed codebook,
given by (11) is always suboptimal and hence leads to performance degradation. The other suboptimality arises from the constraints on the code points in the transformed codebook
in the sense that the Voronoi shape of the transformed code is not matched to the distortion function
, and hence is not optimized to minimize the inertial profile. Note that these two suboptimalities, named as point density loss and cell shape loss, were also discussed in [
22] in the setting of the conventional product quantizers and further applied to study the distortion performance of conventional quantizers with transformed codebooks.
3.3. Characterizing the Inertial Profile of the Transformed Codebook
Unfortunately, the Voronoi region
of the transformed codebook, which is defined to be
is hard to characterize and depends on both the transformation
as well as the distortion function
. In order to characterize the effects of the transformed Voronoi shape on the system distortion, lower and upper bounds of the normalized inertial profile of the transformed code are provided. First, let us consider a suboptimal quantizer
with transformed codebook
that uses a suboptimal encoding process given by
where
is the optimal encoder that is matched to the distortion function
. This suboptimal encoder can be viewed as an extension of the "companding" model introduced by Bennett [
24] to the general vector quantization problem. It was originally used in conventional scalar quantizers, where the encoder is a combination of a monotonically increasing nonlinear mapping
, the compressor, followed by a uniform quantizer; and the corresponding decoder is composed of a uniform decoder followed by an inverse mapping
, the expander. In the case of the generalized vector quantizer discussed here, the Voronoi shape of the suboptimal transformed encoder
can be analytically characterized as
where
is the optimal Voronoi shape of the original codebook
corresponding to distortion function
. Due to the suboptimality of encoder
, the normalized inertial profile of the transformed Voronoi shape
is upper bounded by the inertial profile of
given by (15), but lower bounded by the inertial profile of the optimal Voronoi shape
corresponding to the distortion function
.
Proposition 1.
Under high resolution assumptions, the approximated inertial profile
of a quantizer with transformed codebook can be upper and lower bounded by the following form:
Furthermore, if the source variable is subject to
constraints given by the vector equation
, the constrained inertial profile
can be similarly bounded by
where
is an orthonormal matrix with its columns constituting an orthonormal basis for the null space
.
Proof.
Due to the constraints on the code points in the transformed codebook
, which cannot be optimized to minimize the normalized inertial profile, it is evident that the transformed inertial profile
is lower bounded by the optimal inertial profile
given by (5). Hence, inequality
in (16) can be obtained after some manipulations. The same reasonings are valid for inequality
in (17) for the constrained source.
As for inequality
in (16), since function
is first order continuous, any points in the vicinity of the transformed code point
has a first-order Taylor series expansion given by
Moreover, due to the fact that
is a one-to-one mapping, for any point
in the vicinity of
, there exists a unique point
in the neighborhood of
such that
. Therefore, under high resolutions, the distortion function
can be expanded around point
as follows:
which has quadratic form but with transformed sensitivity matrix. By substituting (19), as well as the Voronoi shape of the suboptimal encoder given by (15), into the definition of the inertial profile given by (3), we can obtain the following normalized inertial profile of the transformed code with suboptimal encoder:
which corresponds to inequality
in (16).
If the source variable (vector)
is further subject to
constraints given by the vector equation
, the distortion function
can be similarly expanded around point
as
where
is the projected error vector with respect to point
given by
By substituting (21) and the suboptimal Voronoi shape (15) into the inertial profile definition (3), we can obtain the suboptimal inertial profile of the transformed code with constrained source
which corresponds to inequality
in (17).
3.4. Distortion Integral of the Transformed Codebook
By substituting the transformed point density (11) and the bounds of the transformed inertial profile given by (16) into the distortion integration (2), we can upper and lower bound the asymptotic system distortion of a transformed quantizer by the following form:
Similarly, by substituting (12) and (17) into (2), the asymptotic distortion of a constrained quantizer with transformed codebook is bounded by
Similar to conventional product transformed quantizers [
22], there exist trade-offs between the two suboptimalities: point density loss and Voronoi shape loss. To be specific, it is always possible to find a transformation
such that the transformed point density
matches exactly the optimal point density
. However, by doing so, the transformation might cause shape loss of the transformed Voronoi cells in some cases, which will lead to significant increase in the normalized inertial profile. Therefore, a transformation that optimally balances two types of losses should be employed. This tradeoff is directly reflected in the distortion bound
where both
and
in (24) depend on the transformation
. So is the distortion bound
given by (25).