WO2009061243A1

WO2009061243A1 - Method and apparatus for ofdm channel estimation in a radio communication system

Info

Publication number: WO2009061243A1
Application number: PCT/SE2007/050828
Authority: WO
Inventors: Shiau-He Tsai; Tsao-Tsen Chen
Original assignee: Telefonaktiebolaget Lm Ericsson (Publ)
Priority date: 2007-11-08
Filing date: 2007-11-08
Publication date: 2009-05-14
Also published as: JP2011504001A; EP2208325A1; JP5228058B2

Abstract

The present invention pertains to radio communications and is concerned with improved channel estimation in relation to reception of OFDMA frames. The OFDM frame includes a preamble (15) and a data portion (17). In a conventional manner, the data portion (17) is provided with pilot symbols according to a predetermined pattern in order to facilitate channel estimation. A correlator (29) correlates a preamble signal of the received OFDM frame with an a-priori known, i.e. known at a receiver (24) side, preamble signal, resulting in a correlator signal z (n). A pilot extractor (35) extracts frequency domain samples corresponding to pilot symbols in the transmitted OFDM frame from the data portion (17) of the received OFDM frame. A channel estimator (39) generates a channel estimate based on the correlator signal z (n) and the extracted frequency domain samples.

Description

Method and apparatus for OFDM channel estimation in a radio communication system.

TECHNICAL FIELD

The present invention pertains to the field of radio communications, and in particular to the part of this field that is concerned with channel estimation.

BACKGROUND

An OFDM (Orthogonal Frequency Division Multiplexing) system makes use of multiple modulated tones (sub-carriers) inside a frequency band intended for digital radio communications. Instead of modulating a time-domain pulse with a data symbol, each sub-carrier is modulated with a data symbol. The modulated sub-carriers are then added together to form a so-called OFDM symbol. As a result, OFDM has the salient feature of being able to process a multipath channel frequency domain distortion on a tone-by-tone basis. Furthermore, since multiple bits are carried in parallel by one OFDM symbol, the symbol duration can be extended by the number of bits in parallel without compromising a data rate.

Other conventional systems either use time domain equalization to mitigate ISI (Inter-Symbol Interference) caused by multipath propagation or use a RAKE receiver to resolve multiple paths and then combine them for better SNR (Signal-to-Noise Ratio) . As bandwidth gets wider, both these techniques have disadvantages in complexity when trying to mitigate the self interface caused by multipath distortion. In the case of time-domain equalization, wider bandwidth translates to shorter data pulse duration and hence more taps are needed in the equalizer. For linear equalizers, introducing more taps will introduce a higher degree of complexity as well as problems with numerical stability of pre-cursor and post-cursor filters. For nonlinear equalization based on sequence detection, the complexity grows exponentially with the number of taps. In the case of the RAKE receiver, wider bandwidth translates into more RAKE fingers, which means that the received signal power of each finger is reduced. The reduced power of each finger may cause performance degradation in acquisition, tracking and channel estimation. The increased number of fingers also implies more correlators and delay lock loops, making receiver hardware much more complicated.

In contrast, the OFDM system utilizes a discrete Fourier transform (DFT) to analyze the multipath effect by an associated discrete frequency transfer function, also referred to as frequency selectivity. Knowledge of the multipath channel is applied in order to compensate for the distortion of each one of the modulated sub-carrier tones. When properly designed, a spacing of the sub-carrier tones is much narrower than a coherence bandwidth of the multipath channel. As a result, the multipath distortion on each sub-carrier tone only manifests itself as a scalar multiplication. The scalar multiplication introduces amplitude change and phase rotation to each modulated sub-carrier tone. With help from channel estimation, this multiplicative scalar distortion simplifies the problem of detecting a data symbol of each carrier tone to "one-shot" detection, as opposed to the deconvolution/sequence detection or maximum ratio combining used in a time-domain equalizer or a RAKE receiver.

In wireless OFDM communications, a reconstruction of the discrete frequency transfer function associated with a multipath radio channel is commonly referred to as channel estimation. Available techniques for channel estimation can be classified to two large categories: one based on pilot insertion and the other one based on Bayesian detection. A great majority of mobile communication standards, if not all, provide pilots to assist coherent demodulation. The theoretical foundation for pilot- based channel estimation is the sampling theorem, which states that in order to reconstruct a band-limited signal from a set of discrete samples, a sampling rate used to sample the signal must be at least twice the highest frequency component of the signal. This requirement is also known as the Nyquist criterion. In an ideal, i.e. noise-free, environment the pilot tones can be seen to provide samples of the multipath channel's discrete frequency transfer function. If the pilot tones are inserted with a frequency spacing that is smaller than one half of the maximum rate of change in the discrete frequency transfer function, the tones between the pilot tones can be perfectly reconstructed, providing there is no additive white Gaussian noise (AWGN) . There are many ways of reconstructing the discrete frequency response of a multipath channel from pilot tones, for example, FIR filtering, interpolation, iteration or singular value decomposition. All these reconstruction methods can be characterized as some type of bandpass filtering. Therefore, no matter what reconstruction method is used, a channel estimation error in practice comes from two sources in a passband: one is irreducible AWGN and the other is a mismatch between the passband filter response and a frequency spectrum of the discrete frequency response pattern. The error from the residual AWGN in the passband is called tracking error, and the one from the mismatch between frequency response pattern and filter response is called model tracking error. A performance of channel estimation is determined by how much these two errors can be suppressed with respect to the power of the reconstructed multipath spectrum.

The channel estimation error can consequently be reduced by setting a correct bandwidth of the reconstruction filter and by selecting the filter response according to a spectral density of the discrete channel response. One known technique is described in L. Deneire, P. Vandenameels, L. Van der Perre, B. Gyselinkckx and M. Engels, "A Low-Complexity ML Channel Estimator for OFDM," IEEE Trans Commun, vol 51, no. 2, Feb 2003, which describes a constraint least square solution (CLS) , which uses a cyclic prefix length as a constraint of the filter bandwidth and digitally removes non-zero points outside of a cyclic prefix window when frequency domain pilot samples are projected onto a time domain during least square interpolation. The CLS solution only makes use of the assumption that no multipath component exists beyond the length of the cyclic prefix to suppress out- of-band noise by setting the pilot-sample IFFT (Inverse Fast Fourier Transform) points outside the cyclic prefix length to zero. After artificially zeroing those points, the channel estimator converts the reduced-noise pilot-sample IFFT point back to the FFT domain. In this case, the constraint is the cyclic prefix length and the operation of setting outside points to zero is equivalent to projecting the signals into a space that is orthogonal to noise-only dimensions. However, there are several drawbacks with this technique. First, the cyclic prefix length, in general, is based on a worst-case consideration, and the multipath delay for an individual case can be much shorter. Setting the constraint to a fixed value misses the possibility of further suppressing out-of-band noise. Second, the projection only addresses the tracking error, not the model tracking error. That is, the CLS method does not consider spectral points that have higher SNR values and put more weights on those points in the channel estimation process. Other existing techniques include predetermined FIR filters and decision-directed adaptive digital filters, where the former is simple but not optimized, and the latter is better for varying conditions but not as numerically robust nor as easy to implement. The generic adaptive filter approach usually performs well when considering the model tracking error but does not perform as well when it comes to suppression of the tracking error.

Consequently there is need for improved channel estimation in relation to OFDM communications which efficiently suppresses both the tracking error and the model tracking error. SUMMARY

One main problem addressed by the present invention is to obtain ways and means for channel estimation relating reception of OFDM frames that provide efficient suppression of the various channel estimation errors.

According to one aspect of the present invention, the above- stated problem is solved with a method for channel estimation relating to reception of an OFDM frame. The OFDM frame includes a preamble and a data portion. In a conventional manner, the data portion has been provided with pilot symbols according to a predetermined pattern in order to facilitate channel estimation. A preamble signal of the received OFDM frame is correlated with an a-priori known (i.e. known at a receiver side) preamble signal, resulting in the generation of a correlator signal. Frequency domain samples corresponding to pilot symbols in the transmitted OFDM frame are extracted from the data portion of the received OFDM frame. A channel estimate is generated based on the correlator signal and the extracted frequency domain samples .

As will be shown in the detailed part of the description, the correlator signal provides an approximation of an impulse response of the multipath radio channel over which the OFDM frame is transmitted. By the duality of the discrete Fourier transform, the correlator signal also provides an approximation of the corresponding discrete transfer function. Admittedly, the information provided by the correlator signal is imperfect

(approximate) . However, the correlator signal nevertheless provides additional input that allows it to serve as valuable model for the channel estimation in combination with the extracted frequency domain (pilot) samples. In particular, the information from the correlator signal provides improved knowledge regarding two important aspects: optimal passband; and frequency weighting. A first source of estimation error comes from noise (AWGN) , which has an unlimited bandwidth. The discrete Fourier transform (DFT) of the discrete frequency response, however, has a limited number of non-zero frequency components, the non-zero signal point with the highest frequency being limited by a maximum delay spread of the mulipath channel. The maximum delay spread, in turn, depends on the constantly changing multipath channel. The information contained in the correlator signal therefore allows for insight into abrupt "birth-death" events in the multipath channel, ensuring a lowest passband for reducing AWGN power and hence minimum tracking errors. A second source of estimation error comes from the fact that the discrete frequency response does not have a constant magnitude within the passband. It follows that, depending on the frequency, some frequency domain (pilot) samples are more important than others. Basing the channel estimation not only on the extracted frequency domain samples but also on the correlator signal consequently improves channel estimation, since the above-mentioned frequency dependence is accounted for through the additional information provided by the correlator signal.

The above-indicated method may, for example, be carried out using a computer program, and the invention includes also such a program as well as a computer program product carrying such a program.

According to another aspect of the present invention, the above- stated problem is solved with a receiver adapted to perform the above-indicated method.

According to yet another aspect of the invention, the above- stated problem is solved with an apparatus capable of performing the above-indicated method.

According to a further aspect of the present invention, the above-stated problem is solved with a receiver including the above-indicated apparatus. According to yet a further aspect of the present invention, the above-stated problem is solved with a radio node or a user equipment including the above-indicated apparatus.

One main advantage of the present invention is that it allows for suppression of both the tracking error and the model tracking error. Furthermore, the invention allows for suppression of these errors in an essentially independent manner. For example, suppression of the tracking error will not negatively influence suppression of the model tracking error.

BRIEF DESCRIPTION OF THE DRAWINGS

Figure 1 is a block diagram illustrating an exemplary cellular network.

Figure 2 is a schematic diagram illustrating a conventional OFDM frame .

Figure 3 is a block diagram illustrating reception of an OFDM frame according to an embodiment of the invention.

Figure 4a is signal diagram illustrating a multipath impulse response .

Figure 4b is a signal diagram illustrating a discrete frequency response corresponding to the impulse response of figure 4a.

Figure 5a is a signal diagram illustrating a spectrum of a signal obtained by downsampling the discrete frequency response of figure 4b.

Figure 5b is signal diagram illustrating a reversal of the signal in figure 5a.

Figure 6 is a block diagram illustrating a pilot filter generator and channel estimator design according to an embodiment of the invention. o

Figures 7a and 7b are signal diagrams illustrating non-uniform to uniform interpolation.

Figure 8 is a block diagram illustrating a pilot filter generator and channel estimator design according to an embodiment of the invention.

Figure 9 is a flow chart illustrating a method for channel estimation according to an embodiment of the invention.

Figure 10 is a flow chart illustrating a method for channel estimation according to an embodiment of the invention.

DETAILED DESCRIPTION

The invention will now be described further using exemplary embodiments and accompanying drawings.

Figure 1 is block diagram illustrating an exemplary cellular network 1, wherein the present invention may be put into use. The cellular network 1 is here, by way of example only, a WiMAX network, but a skilled person will realise that the present invention is applicable to any type of network that employs OFDM or OFDMA techniques. The cellular network 1 includes a base station (BS) 5, which is currently serving a user equipment (UE) 3 via an established radio connection 4. The cellular network includes also another BS 7. A skilled person will realise that the cellular network 1 is simplified in order not to obscure the current presentation with unnecessary detail. For example, a real cellular network will of course include many more base stations. BS 5 and BS 7 are connected to an ASN_GW (Access Service Network Gateway) 9, which in turn is connected to a core network 11 providing, for example, switching and transport functions. The core network 11 is also connected to one or more other networks (not shown) . Through the cellular network 1, the UE 3 may access various communication services, such as radio telephony and data services . In a radio communications environment, such as in the cellular network 1, transmitted radio signals do not normally reach a receiver via a single propagation path. Instead, due to reflections against various objects in the surroundings, the transmitted radio signals will reach the receiver via a multitude of different propagation paths, resulting in a multipath channel. Multipath interference problems may occur if a symbol duration is shorter than a delay spread of an employed multipath channel. The delay spread can be viewed as a maximal difference in delay between the various propagation paths associated with the multipath channel. In contrast to single carrier transmissions, OFDM allows an increase in effective symbol duration by distributing the symbols to multiple sub- carriers. This reduces intersymbol interference (ISI) and increases a probability for interference free transmissions.

A schematic diagram of a conventional OFDM frame structure is shown in figure 2. The OFDM frame in figure 2 is vertically divided into L OFDM symbols, which are here numbered, by way of example, as 0, 1, 2,..., L-I. The frame is further horizontally divided into N sub-carriers, which are here numbered, by way of example, as 0, 1, 2,..., N-I. Each sub-carrier is modulated by one symbol per OFDM symbol. The symbols (one per sub-carrier) of the first OFDM symbol constitute here a preamble 15 of the frame. The preamble may, however, occupy more than one OFDM symbol. The preamble 15 is used, for example, for system acquisition, for sector identification and for frame timing synchronisation. In figure 2, the preamble 15 is followed by a data portion 17, which occupies the remaining OFDM symbols. As is well understood by a person skilled in the art, the data portion 17 may be further subdivided in any number of fields designed for particular types of data, e.g. for signalling data, control data or user data. The OFDM frame structure of figure 2 may also be used to provide multiple access capabilities by assigning different parts of the data portion 17 to different users; this technique is sometimes referred to as OFDMA (Orthogonal Frequency Division Multiple Access) . Furthermore, pilot symbols are usually included in the data portion 17 according to a predetermined pattern in order to support channel estimation, as will be described in greater detail below. A so-called cyclic prefix is also added to each OFDM symbol. The cyclic prefix is a partial repetition of samples of the OFDM symbol. Normally, an end portion of the OFDM symbol is copied to the beginning of the OFDM symbol, before digital-to-analogue conversion is performed. As is well understood by a person skilled in the art, the cyclic prefix is not only introduced to provide a guard period from the preceding OFDM symbol delayed by the multipath, but also to provide point-wise multiplicative mathematical properties associated with the multipath radio channel in the domain of discrete Fourier transform. In particular, the cyclic prefix makes it possible to model an influence of the multipath radio channel as a circular convolution operation. The cyclic prefix reduces intersymbol interference (ISI) and increases a probability for interference free transmissions. Due to the relatively long duration of each OFDM symbol, the overhead of the cyclic prefix is relatively small when averaged over the number of data symbols carried by all the tones.

Figure 3 is a block diagram illustrating reception of OFDM signals according to an embodiment of the invention. An OFDM transmitter 21 transmits an OFDM frame to an OFDM receiver 24 via a multipath radio channel 23. The receiver 24 may, for example, form part of a radio node, such as a BS or a UE. At the receiver 24, the transmitted OFDM frame is received by an antenna as a continuous time signal r(t) . A front end unit 25 is adapted to receive the time signal r(t) . The front end unit 25 performs frequency down-conversion, sampling and digitalisation of the continuous time signal r(t) to produce a corresponding discrete time signal r(n) . A time switch 27 is connected to the front end unit 25 and adapted to receive the discrete time signal r(n) . The time switch 27 divides the discrete time signal r(n) into a first discrete time signal part that corresponds to the preamble of the transmitted OFDM frame and into second discrete time signal part that corresponds to the data portion of the transmitted OFDM frame. The time switch 27 also removes any part of the discrete time signal r(n) that corresponds to a cyclic prefix in the transmitted OFDM frame. The first discrete time signal part thus forms a received preamble signal, which is here denoted w (n) , n e {θ,l, ... , N - l} .

As explained above, the preamble of the transmitted OFDM frame is generated by modulating each sub-carrier with a respective symbol from a predetermined symbol sequence P_k, k e {0,1, ..., N - l} . The sequence P_k can therefore be seen as frequency domain representation of the preamble of the transmitted OFDM frame. The sequence P_k has a corresponding time domain representation in a discrete time preamble signal p(n) , n e {θ,l, ... , N - l} . The frequency domain representation P_k and the corresponding time domain representation p(n) are related through a discrete Fourier transform (DFT) according to

P_k = DFT [p(n)] (k) .

An autocorrelation function R_pp(n) associated with the time domain representation p(n) of the preamble is defined according to

N-I

R_pp(n) = ∑ p(m) • p( (m + n) mod N), n e {θ,l, ... , N - l} , m = 0

where the bar is used to denote complex conjugation. It is convenient to introduce the convention that signal arguments are always interpreted mod N (for an N-point signal); for example, instead of writing (m + n) mod N it then suffices to write m + n, mod N being implicit. This convention will be used in the following, if nothing else is suggested. For example, p(-n) should be interpreted as p(-n mod N) . This convention is of course equivalent to a periodic signal extension. Since the preamble has been selected to be suitable for synchronisation purposes, the time domain representation p(n) of the preamble has strong autocorrelation properties so that R_pp(O) is the clearly dominating value of the autocorrelation function R_pp(n) . It is, therefore, possible to approximate R_pp(n) using a discrete Kronecker delta function δ{n) , so that

^Rp_P(n) R_pp(0) • δ(n) = δ(n),

where Parseval's Theorem has been employed to obtain the last equality. It is easy to show that a correlation operation between two discrete time signals - and therefore also the autocorrelation - can be expressed in terms of the circular convolution operator, hereinafter denoted with the symbol * . For example, the above-mentioned autocorrelation function may be written as

R_pp(n) = p(n) * p(-n) .

The OFDM receiver 24 in figure 3 is provided with a correlator 29 which is adapted to perform a correlation operation between the received preamble signal w(n) and the time domain representation p(n) of the preamble of the transmitted OFDM frame, p(n) being a-priory known to the OFDM receiver 24. A correlator signal z (n) , n e { 0, 1, ..., N-I } , is outputted from the correlator 29. The received preamble signal w(n) is related to the time domain representation p(n) according to

w(n) = V^EsP(ⁿ) * ^h(ⁿ) + Or(Ii) ,

where m(n) represents AWGN, E₃ represents a received signal energy, and where h(n) is a time domain impulse response of the multipath radio channel 23. Consequently, z(n) = R ^N w_wp_p(>n) = w(n) * p(-n) = V^Es (P(ⁿ) * ⁿ(ⁿ) ) * P(^~n) + ^m^ * P(^~n) =

N - I

E_s h(n) * (p(n) * p(-n)) + m(n) * p(-n)

Vk Σ h(n) * δ{n) + C7(n) * p(-n) = O

The fact that the circular convolution operator is commutative, associative and distributive has been used above together with the derived approximation for the autocorrelation function R_pp(n) . The last term involves correlation with a noise component. If this term can be neglected, the correlator signal may be approximated as

Here H_k, k e {θ,l, ••• , N - l} , is the discrete frequency response of the multipath radio channel 23. The discrete frequency response equals the discrete Fourier transform of the impulse response h(n) . The discrete frequency response H_k thus constitutes a frequency domain representation of the impulse response h(n) .

The discrete frequency response H_k is also known as frequency selectivity pattern. From the above, it is clear that the discrete frequency response H_k is essentially proportional to a discrete Fourier transform of the correlator signal z (n) .

The time switch 27 outputs also the above-mentioned second discrete time signal part, which corresponds to the data portion of the transmitted OFDM frame. In figure 3, y(n), n e { 0, 1,..., N- 1}, is used to denote a discrete time signal that constitutes portion of the second discrete time signal part corresponding to one received OFDM symbol of the data portion of the transmitted OFDM frame, e.g. the first OFDM symbol of the data portion. For the sake of simplicity, the operation of the receiver will be described only for the signal y(n) . The other received OFDM symbols are treated in a corresponding manner. The signal y(n) is consequently a discrete time domain representation of one received OFDM symbol. An FFT (Fast Fourier Transform) calculator 33 is connected to the time switch 27 and adapted to receive the signal y(n) . The FFT calculator 33 employs an FFT algorithm to calculate frequency domain samples Y_k, k e { 0, 1, ..., N-I } , corresponding to the signal y(n) . Let X_k, k <≡ { 0, 1, ..., N-I } , denote the symbols transmitted on the sub-carriers of the OFDM symbol associated with the signal y(n) . The frequency domain samples Y_k are related to the transmitted sub-carrier symbols X_k according to

γ_k = V^^"H _k ^• X_k + W_k, k e {0,1, ^••• , N - l},

Where W_k, k e { 0, 1, ..., N-I } , is a frequency domain representation of the AWGN. Note in particular that the influence of the multipath radio channel 23 is here manifested through a purely multiplicative effect. Some of the symbols X_k are a-priori known pilot symbols that have been introduced in the data portion of the OFDM frame in order to facilitate channel estimation. Let K denote a pilot index set of the OFDM symbol associated with the signal y(n), i.e. K CZ { 0, 1, ..., N-I } such that X_k is a pilot symbol if and only if k e K. Note, however, that the pilot index set does not have to be the same for all OFDM symbols in the data portion of the OFDM frame. The receiver 24 includes a pilot extractor 35 and a data extractor 37. The pilot extractor 35 is adapted to extract the frequency domain samples Y_k, k e K, that correspond to the pilot symbols X_k, k e K. The data extractor 37 is adapted to extract the frequency domain samples Y_k, k £ K, that corresponds to actual data symbols X_k, k <£ K, of the OFDM symbol. A channel estimator 39 is connected to the pilot extractor 35. The channel estimator 39 is adapted to generate a channel estimate H_k, k e {θ,l, ••• , N - l}, which estimates the discrete frequency response H_k based on the frequency domain samples Y_k, k e K, extracted by the pilot extractor 35 and the a-priori known pilot symbols X_k, k e K. The channel estimate H_k is outputted from the channel estimator 39 to a demodulator 41, which is adapted to demodulate an output from the data extractor 37 so as to obtain demodulated data. Without loss of generality, the pilot symbols will all be assumed to be equal to one (1) hereinafter.

The receiver of figure 3 further includes a pilot filter generator 31 adapted to receive the correlator signal z (n) . The pilot filter generator is adapted to generate a pilot filter, which is outputted from the pilot filter generator 31 to the channel estimator 39, which uses the pilot filter when processing the output from the pilot extractor 35 in order to obtain the channel estimate H_k , k e { 0, 1, ..., N-I } . The pilot filter is thus used when the channel estimate is generated based on the received frequency domain samples Y_k, k e K, that correspond to the pilot symbols inserted in the transmitted OFDM frame. The pilot filter may, for example, be applied directly to the frequency domain samples Y_k, k e K, to generate the complete channel estimate or a subset of the channel estimate. The pilot filter may also be applied indirectly to the frequency domain samples Y_k, k e K, i.e. after the frequency domain samples have been processed in some manner. Various specific examples of how to apply the pilot filter will be given below.

Figures 4a to 5b provide a simple signal example that illustrates the principles allowing a discrete frequency response to be estimated based on a limited number of samples under ideal circumstances. In effect, these figures may be said to exemplify a discrete counter-part of the sampling theorem. Figure 4a is a signal diagram showing an (multipath) impulse response with 64 taps. However, in this simple example, only tap number zero and tap number two are non-zero (tap zero being equal to 1 and tap two being equal to 0.5) . Figure 4b is diagram illustrating a discrete frequency response corresponding to the impulse response of figure 4a. Figure 4b only shows magnitude (absolute value) , thus leaving phase information aside. It is now assumed that a new signal is obtained by downsampling the discrete frequency response, i.e. by collecting the values of the discrete frequency response at regular intervals. For example, assume that every fourth value of the discrete frequency response is collected, then a downsampled signal with sixteen points will be obtained from the original 64-point discrete frequency response. Figure 5a shows a spectrum (DFT) of this downsampled signal. Figure 5b is a diagram that illustrates a reversal of the spectrum of figure 5a. Here the term reversal is used such that if f (n) denotes an arbitrary discrete signal, then the reversal of this signal is the signal f (-n) . As can be seen, the signal of figure 5b is directly proportional to the original impulse response of figure 4a (but limited to the first sixteen points) . This is no accident. To understand why, one needs to know two things: firstly, how the spectrum (DFT) of the discrete frequency response is related to the impulse response; and secondly, how a spectrum of a downsampled signal is related to a spectrum of the signal from which the downsampled signal is obtained. It is first noted that the discrete Fourier transform DFT may be expressed in terms of the inverse discrete Fourier transform DFT^"1, i.e. according to DFT [f (n) ] (m) = N-DFT^"1 [f (n) ] (-m) , where again f (n) is an arbitrary (N-point) discrete signal. Since the discrete frequency response of figure 4b is the discrete Fourier transform of the impulse response of figure 4a, it immediately follows that the spectrum of discrete frequency response is proportional to a reversal of the impulse response. Furthermore, when a signal is generated by downsampling an original signal, the spectrum of the downsampled signal can be obtained from the spectrum of the original signal by partitioning this spectrum into a number of portions of equal length, adding the portions together and multiplying the result with a proportionality factor. The number of portions and the proportionality factor both depend on the rate at which the downsampling is performed. In this signal example, the downsampling rate is such that every fourth signal point of the discrete frequency response is sampled. The downsampled signal therefore has sixteen points (64/4) . It follows that the impulse response can be unambiguously (assuming no other disturbances) constructed from the downsampled version of the discrete frequency response, as long as the non-zero taps of the impulse response of figure 4a are limited to the first sixteenth taps. Consequently, it also follows that the (complete) discrete frequency response may be reconstructed from the downsampled version of the frequency response.

It may be noted that the delay spread of a multipath radio channel is strictly limited, normally less than a designated cyclic prefix length. Therefore h(n) = 0, for n > n₀, where n₀ denotes a maximum delay spread index. It may further be noted that the cyclic prefix length and the pilot insertion rate are inter-related. If the pilot insertion rate does not exceed the cyclic prefix percentage of an OFDM symbol, then the frequency selectivity pattern cannot be reconstructed from the pilot samples without aliasing. Thus an insufficient pilot insertion rate will result in inter-tone interference in channel estimation even without any noise.

Figure 6 is block diagram illustrating according to one embodiment a design of the pilot filter generator 31 and the channel estimator 39 that takes advantage of the above- exemplified principles. In the embodiment of figure 6, the pilot filter generator 31 includes a squared magnitude unit 51, which receives the correlator signal z (n) and calculates a squared magnitude of the correlator signal z (n) . The output from the unit 51 is a squared magnitude signal |z(n)| , n e { 0, 1, ..., N-I } , which is provided to a pilot filter synthesizer 53. The pilot filter synthesizer 53 is adapted to generate a pilot filter in the form of a FIR (Finite Impulse Response) filter f_k, k e { 0, 1, ..., K-I } . Optionally, the pilot filter generator 39 of figure 6 may include also a filter which filters the correlator signal z(n) before it is delivered to the unit 51. Such a filter may be provided in order to suppress self-interference and noise for more accurate estimation.

As described previously, the correlator signal z (n) provides an approximation to the discrete Fourier transform of the discrete frequency response, which is to be estimated. This means that a suitable passband for the pilot filter can be obtained through the information provided by correlator signal z(n) . In particular, the FIR filter may be determined such that its passband is based on the magnitude (absolute value) of the correlator signal.

One effective selection for the FIR filter would be for its passband only to have non-zero values at points at which the magnitude of the correlator signal is non-negligible, e.g. greater than a predetermined value. The reason for this selection is that a negligible signal value of the magnitude of the correlator signal is related to a weak spectral component of the discrete frequency response, and letting such a spectral component pass only contributes noise and provides in principle no information that can contribute to the channel estimation.

Another effective selection of the FIR filter f_k would be if it fulfilled the following condition

DFT -1 [f_{k *} f__kJn) - |h(nf, for n = 0,1,...,K-I

In this case, the passband of the FIR filter is such that it puts a higher weight on spectral components of the discrete frequency response having a higher SNR (Signal to Noise Ratio) and a lower weight on spectral components having a lower SNR. For spectral component that consist mainly of noise (very low SNR) , the FIR filter notches them out, thereby reducing also the tracking error.

Yet another effective selection of the FIR filter f_k can be found by applying an MMSE (Minimum Mean Squared Error) criterion, which leads to the following condition for the FIR filter , for n = 0,1,...,K-I,

where σ_m ², denotes a variance of CJ' (n) (indicating noise and interference) . When the variance is small, the FIR filter is equivalent to a bandpass filter having a unit-gain passband that falls on non-negligible spectral components of the discrete frequency response. When the variance is large, a shape of the passband of the FIR filter approaches the shape of an energy spectrum of the discrete frequency response. The MMSE criterion thus strikes an appropriate balance between noise reduction and distortion of the multipath radio channel.

The conditions above are formulated in terms of the impulse response h(n), which in principle is unknown. However, as was shown earlier, the correlator signal z(n) is (approximately) related to the impulse response h(n) . According to particular embodiments of the invention it is therefore suggested to replace |h(n)| with C^~ • |z(n)| in the above conditions for selecting the FIR filter f_k. Consequently, according to these embodiments, the FIR filter is generated such that an inverse discrete Fourier transform of an autocorrelation of the FIR filter obtains a predetermined relation to the squared magnitude of the correlator signal z (n) . As is well understood by a person skilled in the art, this means that a power spectrum of the FIR filter will be (approximately) based on a power spectrum of the discrete frequency response. The channel estimator 39 of figure 6 is adapted to generate the channel estimate H_k by a linear filtering operation involving the FIR filter f_k and the output from the pilot extractor 35. However, the basic filtering operation in figure 6 works in principle only when the members of the pilot index set K are evenly distributed. Such uniformly distributed pilots are often referred to as comb pilots. This means that the pilot frequency indices k are integers which are evenly distributed so that for every N_comb sub-carriers there is one sub-carrier carrying a pilot. Here N_comb is an integer value indicating a frequency index separation between sub-carriers carrying pilots. Consequently, a pilot index set K may be said to define a "comb" if an only if (VkI, k2 e K) (N_comb divides kl - k2) . If K denotes the number of elements in the pilot index set, then K=N/N_comb, where it is assumed that N_comb is a divisor of N.

The embodiment of figure 6, however, overcomes the above- described difficulty through the provision of a non-uniform to uniform interpolator 55. If the pilot index set K does not define a comb, the non-uniform to uniform interpolator 55 interpolates between the frequency domain samples Y_k, k e K, to obtain evenly distributed frequency domain samples Y_k , k e K' , where K' is a modified pilot index set that defines a comb. The principle is illustrated in figures 6a and 6b. Figure 6a illustrates a signal diagram with non-uniformly distributed frequency domain samples (indicated by vertical lines with bullets at the top) and a curve obtained by interpolating between the non-uniformly distributed samples. In figure 6a, the frequency domain samples are provided as a function of a digital frequency instead of the frequency index number k. The digital frequency is simply the frequency index number k divided by N. Figure 6b is a diagram that illustrates how the interpolation curve of figure 6a can be used to generate uniformly distributed frequency domain samples. Standard generic non-uniform to uniform interpolation techniques may be used by the interpolator 55, e.g. as described in F. Marvasti, "Non-uniform Sampling: Theory and Practice", chapters 3 and 4, Springer, 1^st edition 2001. This type of interpolation does not make use of a-priory information relating to the multipath radio channel 23 and is usually equivalent to an all-pass filter.

Turning again to figure 6, the channel estimation will be described assuming that the pilot index set K is consistent with a comb such that K = {mN_comb: m e { 0, 1, ..., K-I } } . If not, the non- uniform to uniform interpolator 55 is used to generate the evenly distributed frequency domain samples Y_k , k e K' , and the procedure will be the same but using the samples Y_k and the modified pilot index set K' (K' = {mN_comb: m e { 0, 1, ..., K-I } } ) instead of the samples Y_k and the original pilot index set K. The channel estimator 39 in figure 6 includes a circular convolution unit 57, which is adapted to calculate values H_k , k e K, of the channel estimate by performing a circular convolution between the FIR filter f_k and the frequency domain samples provided by the pilot extractor 35, e.g. in accordance with the following equations

K-I

^l-Ncomb ⁼ 2-i ^Ym-Ncomb ' %-m)modK ' f°^r 1 ^e { 0 , 1 , ..., K-I } . m = 0

The circular convolution unit 57 can thus be seen as generating a subset of the channel estimate. The output from the circular convolution unit 57 may be filtered by a smoothing filter 59 before being provided to a reconstruction interpolator 61. The reconstruction interpolator generates the remaining values H_k , k £ K, of the channel estimate by interpolating between the values of the channel estimate provided by the circular convolution unit 57. An output from the reconstruction interpolator 61 is thus the complete channel estimate H_k , k e { 0, 1, ..., N-I } , which is then provided to the data demodulator 41.

The blocks of figure 6 may be implemented using application specific circuitry (ASIC) or programmable circuitry, or using any combination thereof. Parts of or all of the functionality of the blocks in figure 6 may also be implemented by a computer programmed with suitable computer software.

According to another embodiment, an MMSE technique may again be employed to generate the pilot filter. For this purpose, it is convenient to introduce a matrix formalism. Let T_NxN denote an

FFT matrix with N rows and N columns, where the rows and columns are numbered from 0 to N-I. The FFT matrix is a matrix representation of the discrete Fourier transform. As is well understood by a person skilled in the art, the element of row 1 and column m of T_NxN is thus given by _e ^~2πi^lm/N _? where j is the imaginary unit. Furthermore, let T^_xN denote a modified FFT matrix obtained by removing those rows of T_NxN for which the corresponding row number does not appear in the pilot index set K. The number of rows in the modified FFT matrix is consequently K, i.e. the number of element in the pilot index set K. Let H = [H₀ H₁ ... H_N-1J be a vector representation of the channel estimate, which may be expressed according to the following equation

H = F_NxK • T_KxN • z ,

where z = [z(0) z(l) ••• z(N - l)]^τ and F_NxK is a filter matrix having N rows and K columns. The modified FFT matrix T_KxN has a two-fold purpose in this equation. Firstly, the modified FFT matrix performs a conversion into a frequency domain representation. Secondly, the modified FFT matrix performs a restriction to frequency components having frequency indices in the pilot index set K. In the above-equation, the filter matrix is still an unknown quantity which has to be selected in an "intelligent" manner. The goal is to minimize an estimation error

p ≡ (h - T_N ⁺ _XN • Hj (h -T_N ⁺ _xNHJ.

Here, the plus sign + as a superscript is used to denote the Hermitian conjugate of a vector or a matrix, i.e. the complex conjugation of the transpose of the vector or matrix. Using the equations for the correlator signal obtained earlier, the estimation error can be written as

p = [h - A_NxN • (Ch + m' )]⁺[h - A_NxN • (Ch + m')]_r

where

^ANxN - ¹NxN ^rNxK ¹KxN '

h = [h(0) h(l) ••^■ h(N - l)]^τ, and

m' = [m' (0) m' (1) ••• m' (N - l)]^τ .

According to a Wiener solution, an optimal filter matrix may be obtained according to

F_Nχ_κ = τ_NxN ^• E{: • h • h⁺} • E{C² • h ■ h⁺ + R₅₇.^)^"1 • (O⁺,

where R_8Ji₈₇! denotes a noise covariance matrix and E is used to denote a statistical average (expected value) . The center part of the equation for the filter matrix, i.e. the part that involves the two statistical averages, may be approximated by replacing averaging over the outer product h • h⁺ with averaging over the outer product z • Z⁺ . For example, the following approximation may be used c{h • h⁺ - l 1 - λ

C

1 - λ^m i Σ < m*^'

In the above expression, zi = z for i=m (m being a current OFDM frame number) ; for 0 < i < m, z_± denotes corresponding vector representations of correlator signals generated from preambles of previously received OFDM frames. The average is a weighted average, where the weights are determined by a memory parameter λ (λ < 1) . Furthermore, the noise covariance matrix may be approximated by an identity matrix scaled by an estimated sample variance of noise from an automatic gain control or other noise estimator. The inverse matrix EJC² • h • h⁺ + R_m,_m,j may then be simplified, e.g. by using the Sherman-Morrison formula, which may be expressed according to the following. Suppose A is an invertible square matrix and u, v are vectors.

Suppose further that l + u⁺A^~'v≠0. Then

(_{A +}V¹ A^-1Uv⁺A^"1 IA + uv ) = 7— .

^V ' 1 + U⁺A-¹V

Since isjh-h⁺j is updated one preamble at the time, the matrix

E₎C • h • h⁺ +

can be obtained recursively, provided R_WSJ ^~ has been made available.

Note that the dimension of the projection matrix E{c-h-h⁺}-E{c h h⁺ +R_mtt,}^~ can be reduced to the maximum delay spread detected by the correlator signal z. For example, if the correlator signal shows that the multipath radio channel only exhibits negligible power after a certain number of taps R, the vector T_KxN-z should be truncated to only the first R elements and the projection matrix E)C -h -h j-iijc •h-h⁺+R_S7β7| should only be done with a dimension of R x R. The reduced-dimension projection matrix can be obtained by averaging the outer product of truncated z_± vectors. After the reduced-dimension projection is done, the vector of length R is padded with zeros to length N and transformed back to the frequency domain.

Once the optimal filter matrix has been calculated, the channel estimate can be finally calculated from the frequency domain samples Y_k, k e K, and this calculation may be expressed as a matrix multiplication according to

H = F_NxK • Y ,

where Y = [Y_k(0) Y_k(1) •••

• The index values k (m) m { 0, 1, ..., K-I } , are the elements of the pilot index set organized in increasing order.

Figure 8 is a block diagram illustrating a design according to one embodiment by which the pilot filter generator 31 and the channel estimator 39 are allowed to carry out the just described process for channel estimation. The block diagram in figure 8 includes an FFT matrix generator 71 adapted to generate the FFT matrix T_NxN . An FFT matrix modifier 73 is connected to the FFT matrix generator 71. The matrix modifier

73 is adapted to generate the modified FFT matrix T_KxN based on the FFT matrix T_NxN provided by the FFT matrix generator 71.

The pilot filter generator 31 in figure 8 includes an outer product generator 75, which is adapted to calculate the outer product zz⁺ associated with the correlator signal z (n) . An OFDM frame averaging unit 76 is connected to the outer product generator 75 and adapted to receive an output from the outer product generator 75. The OFDM frame averaging unit 76 calculates the above-mentioned weighted average based on the output from the outer product generator 75 and stored corresponding outer product data relating to previously received OFDM frames. The pilot filter generator 31 further includes a filter matrix generator 77, which is connected to the OFDM frame averaging unit 76, to the FFT matrix generator 71 and to the FFT matrix modifier 73. The filter matrix generator 77 is adapted to generate the filter matrix F_NxK , as described above. In figure 8, the channel estimator 39 includes a matrix multiplier 79, which is connected to the filter matrix generator 77 for receiving the filter matrix F_NxK . The matrix multiplier 79 is further connected to the pilot extractor 35 for receiving the frequency domain samples Y_k, k e K, which are associated with the pilot symbols introduced in the OFDM symbol currently being processed. The matrix multiplier 79 is adapted to generate the channel estimate vector H through a matrix multiplication involving the filter matrix F_NxK and the vector Y, as described above. The channel estimate is then provided to the demodulator 41. I may be noted that this technique has the advantage that it works for uniform pilot insertion as well as for non-uniform insertion.

The filter matrix F_NxK consists of three parts (factors) . The first part is the matrix T_KxN ⁺ that performs a transformation from the frequency domain to the time domain. The second part is a projection matrix, £JC-h-h⁺j- EjC •h-h⁺+R_raa7| , that performs a projection into a linear (sub) space spanned by the multipath response vectors observed from the preamble correlator signals, while considering noise-plus-interference covariance according to the MMSΕ criterion. The third part is the FFT matrix, T_NxN , which is responsible for a transformation back to the entire OFDM spectrum.

The blocks of figure 8 may be implemented using application specific circuitry (ASIC) or programmable circuitry, or using any combination thereof. Parts of or all of the functionality of the blocks in figure 8 may also be implemented by a computer programmed with suitable computer software. Figure 9 is a flow chart illustrating a method for channel estimation relating to reception of OFDM frames according to an embodiment of the invention. Since the method of figure 9 corresponds to the operation of previously described embodiments (such as the embodiments described and indicated in connection with figure 6) the methodology will only be briefly discussed in order to keep the present presentation concise. At a block 91 an OFDM frame is received. Following the receipt of the OFDM frame, a correlator signal is generated at a block 93. The correlator signal is generated by correlating a received preamble signal of the received OFDM frame with an, at the receiver, a-priory known preamble signal corresponding to the actually transmitted preamble. After generation of the correlator signal, a squared magnitude of the correlator signal is generated at a block 95. At a block 97 a pilot filter in the form of a FIR filter is generated based on the generated squared magnitude, e.g. by applying the previously described techniques. At a block 99, frequency domain samples corresponding to pilot symbols of the transmitted OFMD frame are extracted from a data portion of the received OFDM frame. A determination is then made at a block 101 as to whether the extracted frequency domain samples are evenly distributed in the frequency domain, i.e. the pilot index set defines a "comb". If the frequency domain samples are evenly distributed, a circular convolution is performed between the frequency domain samples and the generated FIR filter at a block 105. In case the frequency domain samples are not evenly distributed, a set of evenly distributed frequency domain samples is obtained at a block 103 by applying non-uniform to uniform interpolation to the original frequency domain samples; and the circular convolution at the block 105 is then performed with the set of evenly distributed frequency domain samples rather than with the originally extracted frequency domain samples. In either case, the circular convolution at the block 105 generates a subset of the channel estimate, as explained earlier. After application of a smoothing filter (optional) to this subset at a block 107, the entire channel estimate is obtained from the subset by application of reconstruction interpolation at a block 109.

Figure 10 is a flow chart illustrating a method for channel estimation relating to reception of OFDM frames according to another embodiment of the invention. Since the method of figure 10 corresponds to the operation of previously described embodiments (such as the embodiments described and indicated in connection with figure 8) the methodology will only be briefly discussed in order to keep the present presentation concise. At a block 111, an OFDM frame is received. Following the receipt of the OFDM frame, a correlator signal is generated at a block 113. As before, the correlator signal is generated by correlating a received preamble signal of the received OFDM frame with an, at the receiver, a-priory known preamble signal corresponding to the actually transmitted preamble. At a block 115 an outer product is generated based on the correlator signal, as described earlier. After generation of the outer product, averaging (as described earlier) with corresponding outer products generated from previously received OFMD frames is performed at a block 117. At a block 119, the FFT matrix is generated; and at a block 121, the modified FFT matrix is generated based on the FFT matrix. At a block 123, the filter matrix is generated based on the result of the OFDM frame outer product averaging and using the FFT matrix and the modified FFT matrix, as described earlier. After extracting frequency domain samples corresponding to pilot symbols included in the transmitted OFDM frame at a block 125, the channel estimate is generated at a block 127 by multiplying a vector made up of the extracted frequency domain samples with the generated filter matrix.

Above, the invention has been described using various embodiments. These embodiments are, however, intended only as non-limiting examples, and the scope of protection afforded is instead defined by the appending claims.

Claims

1. A method for channel estimation relating to reception of an OFDM frame which has been transmitted over a multipath radio channel (23) and which includes a preamble (15) and a data portion (17), the method characterised by: generating (93; 113) a correlator signal by correlating a received preamble signal with an a-priori known preamble signal; extracting (99; 125) a first set of frequency domain samples from a received data portion signal, the first set of frequency domain samples corresponding to pilot symbols included in the data portion of the transmitted OFDM frame; and generating (103,105,107,109; 127) a channel estimate based on the correlator signal and the first set of frequency domain samples.

2. The method according to claim 1, wherein: the method further comprises generating (97; 123) a pilot filter in dependence of the correlator signal; and wherein the step of generating the channel estimate involves generating the channel estimate based on the first set of frequency domain samples and using the pilot filter.

3. The method according to claim 2, wherein the step of generating (97; 123) the pilot filter involves generating a linear pilot filter.

4. The method according to claim 3, wherein the step of generating (93) the linear pilot filter involves generating a finite impulse response filter.

5. The method according to claim 4, wherein the step of generating (97) the finite impulse response filter involves generating the finite impulse response filter such that a passband of the finite impulse response filter is determined based on a magnitude of the correlator signal.

6. The method according to claim 5, wherein the inverse discrete Fourier transform of the autocorrelation of the finite impulse response filter is proportional to the squared magnitude of the correlator signal.

7. The method according to claim 5, wherein the inverse discrete Fourier transform of the autocorrelation of the finite impulse response filter equals a fractional expression involving the squared magnitude of the correlator signal.

8. The method according to any one of the claims 4 to 7, wherein the step of generating the channel estimate includes the steps of: determining (101) whether the frequency domain samples in the first set of frequency domain samples are evenly distributed in the frequency domain; if the frequency domain samples in the first set of frequency domain samples are evenly distributed in the frequency domain, performing the steps of: generating (105) a subset of the channel estimate by performing a circular convolution operation between the first set of frequency domain samples and the finite impulse response filter; and generating (109) a complete channel estimate by applying interpolation to the subset of the channel estimate; if the frequency domain samples in the first set of frequency domain samples are not evenly distributed in the frequency domain, performing the steps of: generating (103) a set of evenly distributed frequency domain samples by applying non-uniform to uniform interpolation to the first set of frequency domain samples; generating (105) a subset of the channel estimate by performing a circular convolution operation between the set of evenly distributed frequency domain samples and the finite impulse response filter; and generating (109) a complete channel estimate by applying interpolation to the subset of the channel estimate .

9. The method according to claim 3, wherein the step of generating the linear filter involves generating a filter matrix that may be factored into a first, a second and third part, the first part performing transformation from frequency domain to time domain, the second part being a projection matrix performing a projection into a subspace generated by a vector representation of the correlator signal and a number of corresponding vector representations of correlator signals obtained from previously received OFDM frames, and the third part performing transformation from the time domain to an entire spectrum of the channel estimate.

10. The method according to claim 9, wherein the projection matrix is generated based on an OFDM frame averaged outer product of the vector representation of the correlator signal and an estimated noise covariance matrix.

11. The method according to claim 9 or 10, wherein the step of generating the channel estimate involves performing (127) a matrix multiplication between the filter matrix and a vector made up of the frequency domain samples in the first set of frequency domain samples.

12. The method according to any one of the claims 1 to 11, wherein the method further comprises: extracting a second set of frequency domain samples from the received data portion; and demodulating the second set of frequency domain samples using the generated channel estimate.

13. The method according to claim 12, wherein the second set of frequency domain samples corresponds to actual data included in the data portion.

14. A computer program, characterised by comprising program code for performing a method according to any one the claims 1 to 13.

15. A computer program product, characterised by comprising a computer program according to claim 14.

16. A radio receiver (24) for receiving OFDM frames, the radio receiver characterised by being adapted to perform a method according to any one of the claims 1 to 13.

17. The radio receiver (24) according to claim 16, wherein the radio receiver forms part of a radio communication node for a radio communications system (1) .

18. The radio receiver according to claim 16, wherein the radio receiver forms part of a wireless user equipment.

19. An apparatus (24) for channel estimation relating to reception of an OFDM frame which has been transmitted over a multipath radio channel and which includes a preamble (15) and a data portion (17), the apparatus being characterised in that it comprises : means (29) for generating a correlator signal by correlating a received preamble signal with an a-priori known preamble signal; means (35) for extracting a first set of frequency domain samples from a received data portion signal, the first set of frequency domain samples corresponding to pilot symbols included in the data portion of the transmitted OFDM frame; and means (39) for generating a channel estimate based on the correlator signal and the first set of frequency domain samples.

20. The apparatus according to claim 19, wherein: the apparatus further comprises means (31) for generating a pilot filter in dependence of the correlator signal; and wherein the means (39) for generating the channel estimate are adapted for generating the channel estimate based on the first set of frequency domain samples and using the pilot filter.

21. The apparatus according to claim 20, wherein the means (31) for generating the pilot filter are adapted to generate a linear pilot filter.

22. The apparatus according to claim 21, wherein the means (31) for generating the pilot filter are adapted to generate a finite impulse response filter.

23. The apparatus according to claim 22, wherein the means (31) for generating the pilot filter are adapted to generate the finite impulse response filter such that a passband of the finite impulse response filter is determined based on a magnitude of the correlator signal.

24. The method according to claim 23, wherein the inverse discrete Fourier transform of the autocorrelation of the finite impulse response filter is proportional to the squared magnitude of the correlator signal.

25. The apparatus according to claim 23, wherein the inverse discrete Fourier transform of the autocorrelation of the finite impulse response filter equals a fractional expression involving the squared magnitude of the correlator signal.

26. The apparatus according to any one of the claims 22 to 25, wherein the means (39) for generating the channel estimate includes : means (55) for generating a set of evenly distributed frequency domain samples by applying non-uniform to uniform interpolation to the first set of frequency domain samples in case the first set of frequency domain samples are not evenly distributed in the frequency domain; means (57) for generating a subset of the channel estimate by performing circular convolution between the pilot filter and the first set of frequency domain samples in case the first set of frequency domain samples are evenly distributed in the frequency domain and between the pilot filter and the generated set of evenly distributed frequency domain samples in case the first set of frequency domain samples are not evenly distributed in the frequency domain; and means (61) for reconstructing the entire channel estimate based on the subset of the channel estimate.

27. The apparatus according to claim 21, wherein the means (31) for generating the pilot filter are adapted to generate a filter matrix that may be factored into a first, a second and third part, the first part performing transformation from frequency domain to time domain, the second part being a projection matrix performing a projection into a subspace generated by a vector representation of the correlator signal and a number of corresponding vector representations of correlator signals obtained from previously received OFDM frames, and the third part performing transformation from the time domain to an entire spectrum of the channel estimate.

28. The apparatus according to claim 27, wherein the means (31) for generating the pilot filter are adapted to generate the projection matrix based on an OFDM frame averaged outer product of the vector representation of the correlator signal and an estimated noise covariance matrix.

29. The apparatus according to claim 27, wherein means for generating the channel estimate (39,79) are adapted to generate the channel estimate by performing a matrix multiplication between the filter matrix and a vector made up of the frequency domain samples of the first set of frequency domain samples.

30. The apparatus according to any one of the claims 19 to 29, wherein the apparatus further comprises: means (37) for extracting a second set of frequency domain samples from the received data portion; and means (41) for demodulating the second set of frequency domain samples using the generated channel estimate.

31. The apparatus according to claim 30, wherein the second set of frequency domain samples corresponds to actual data included in the data portion (17) .

32. A radio receiver for receiving OFDM frames, the radio receiver being characterised by comprising an apparatus according to any one of the claims 19 to 31.

33. A radio node for a radio communications system, the radio node being characterised by comprising an apparatus according to any one of the claims 19 to 31.

34. A user equipment for communication with a radio communications system, the user equipment being characterised by comprising an apparatus according to any one of the claims 19 to 31.

35. An apparatus for channel estimation relating to reception of an OFDM frame which has been transmitted over a multipath radio channel (23) and which includes a preamble (15) and a data portion (17), the apparatus being characterised in that it comprises : a correlator (29) operable to generate a correlator signal by correlating a received preamble signal with an a-priori known preamble signal; a pilot extractor (35) operable to extract a first set of frequency domain samples from a received data portion (17) signal, the first set of frequency domain samples corresponding to pilot symbols included in the data portion of the transmitted

OFDM frame; and a channel estimator (39) operable to generate a channel estimate based on the correlator signal and the first set of frequency domain samples.

36. The apparatus according to claim 35, wherein: the apparatus further comprises a pilot filter generator (31) operable to generate a pilot filter in dependence of the correlator signal; and wherein the channel estimator (39) is operable to generate the channel estimate based on the first set of frequency domain samples and using the pilot filter.

37. The apparatus according to claim 36, wherein the pilot filter generator (31) is operable to generate a linear pilot filter.

38. The apparatus according to claim 37, wherein the pilot filter generator (31) is operable to generate the pilot filter as a finite impulse response filter. o o

39. The apparatus according to claim 38, wherein the pilot filter generator (31) is operable to generate the finite impulse response filter such that a passband of the finite impulse response filter is determined based on a magnitude of the correlator signal.

40. The method according to claim 39, wherein the inverse discrete Fourier transform of the autocorrelation of the finite impulse response is proportional to the squared magnitude of the correlator signal.

41. The apparatus according to claim 39, wherein the inverse discrete Fourier transform of the autocorrelation of the finite impulse response equals a fractional expression involving the squared magnitude of the correlator signal.

42. The apparatus according to any one of the claims 38 to 41, wherein the channel estimator (39) includes: an non-uniform to uniform interpolator (55) operable to generate a set of evenly distributed frequency domain samples by applying non-uniform to uniform interpolation to the first set of frequency domain samples in case the first set of frequency domain samples are not evenly distributed in the frequency domain; a circular convolution unit (57) operable to generate a subset of the channel estimate by performing circular convolution between the pilot filter and first set of frequency domain samples in case the first set of frequency domain samples are evenly distributed in the frequency domain and between the pilot filter and the generated set of evenly distributed frequency domain samples in case the first set of frequency domain samples are not evenly distributed in the frequency domain; and a reconstruction interpolator (61) operable to reconstruct the entire channel estimate based on the subset of the channel estimate .

43. The apparatus according to claim 37, wherein the pilot filter generator (31) is operable to generate a filter matrix that may be factored into a first, a second and third part, the first part performing transformation from frequency domain to time domain, the second part being a projection matrix performing a projection into a subspace generated by a vector representation of the correlator signal and a number of corresponding vector representations of correlator signals obtained from previously received OFDM frames, and the third part performing transformation from the time domain to an entire spectrum of the channel estimate.

44. The apparatus according to claim 43, wherein the pilot filter generator (31) is operable to generate the projection matrix based on an OFDM frame averaged outer product of the vector representation of the correlator signal and an estimated noise covariance matrix.

45. The apparatus according to claim 44, the channel estimator (39) is operable to generate the channel estimate by performing a matrix multiplication between the filter matrix and a vector made up of the frequency domain samples of the first set of frequency domain samples.

46. The apparatus according to any one of the claims 35 to 45, wherein the apparatus further comprises: a data extractor (37) operable to extract a second set of frequency domain samples from the received data portion; and a demodulator (41) operable to demodulate the second set of frequency domain samples using the generated channel estimate.

47. The apparatus according to claim 46, wherein the second set of frequency domain samples corresponds to actual data included in the data portion.

48. A radio receiver for receiving OFDM frames, the radio receiver being characterised by comprising an apparatus according to any one of the claims 35 to 47.

49. A radio node for a radio communications system, the radio node being characterised by comprising an apparatus according to any one of the claims 35 to 47.

50. A user equipment for communication with a radio communica- tions system, the user equipment being characterised by comprising an apparatus according to any one of the claims 35 to 47.