The perils involved when pushing direct conversion to the Nyquist limit

Direct conversion receivers often use dual matched active filters. These are generally limited to about 20MHz and have a usable bandwidth of some 40MHz, but apply high sample rates relative to the signal bandwidth. Increasing the filter bandwidth to the Nyquist boundary requires an aggressive filter with very little transition band if receiver selectivity is defined by the low pass (LP) filters. A practical software defined receiver – single or multichannel, tuned largely by the local oscillator (LO) and intended to produce a few MHz of usable bandwidth after image rejection processing – may use a 10MHz dual matched active filter and a sample rate of 100Msample/s. Often, this may be the only filtering required but, because LP filters are generally low order, a high oversampling ratio is needed for true antialiasing. While antialiasing filtering can be relaxed if out of band power is suppressed prior to the mixer, this ignores the bandwidth restrictions that may be required to prevent overloading the low noise amplifier or mixer with out of band interferers and any additional LO/rf suppression that may be required by an active filter. For example, the LT6604-10, a fourth order dual filter/driver, has approximately 70dB of attenuation at 90 to 100MHz, adequate for most applications. The use of SAW filters in the rf domain may mitigate the suppression needs in alias bands, while an integrated LP filter in the demodulator may alleviate out of band filtering requirements, but these are often in the 250 to 400MHz range and only effective in suppressing LO or rf feed through. Active LP filters should have additional suppression beyond 50MHz to handle noise from the driver section that may persist beyond the transition band. Gain and phase mismatch limit the achievable spurious free dynamic range (SFDR) in direct conversion, due to image rejection limitations. This is often seen as simply a case of correcting for gain and phase. For a high degree of image rejection, the mismatch in pass band ripple and delay over the band of interest must also be corrected. Because of tolerance variations, differential filters may display highly localised features in passband phase and amplitude response, hence require more complex correction. Simple time domain correction may become unmanageable approaching the edge of the passband of a high order LP filter. Attempts to use all the Nyquist band do not leave open the option of using only a fraction of the analogue filter's passband. If only half of the potential direct down conversion bandwidth were of interest – say only the positive frequencies (+1) – image rejection of the second Nyquist zone (+2) and the images in the first Nyquist zone (-1) is possible, both zones being negative frequencies. This requires a filter prior to the mixer to suppress those frequencies that would land in the image of the second Nyquist zone (-2, positive frequencies), as well as suppressing the third Nyquist zone above LO. This requires SAW filters prior to the mixer with a nominal bandwidth of 1.5x the sample frequency, centred at 0.25fs from the LO. The availability of high speed a/d converters with an SFDR of 100dB, such as the 16bit 130Msample/s LTC2208, implies that operation is possible in the presence of very strong interferers, but image rejection approaching this order requires extraordinary measures. The desire for wide bandwidth in direct conversion is understandable, since image suppression processing with quadrature signals potentially doubles the bandwidth of an IF sampling receiver at a given sample rate. But wide bandwidth and low passband ripple in an IF sampling (undersampling) scenario requires a high centre frequency. This, in turn, limits dynamic range in many a/d converters and amplifiers. The practical IF filter passband is arguably limited to 20% of the centre frequency. The increasingly common target of 100MHz of usable bandwidth would suggest a 500MHz IF and more than 200Msample/s, which implies high power consumption. It is reasonable to expect direct down conversion to require less power, as amplifiers suitable for baseband frequencies will require less power than high IF amplifiers. And high IF sampling will require repeated amplification, as IF filter insertion loss is much higher than a LP filter. A high degree of selectivity often requires cascaded filters. In IF sampling, two stages of SAW filters – required for some 80dB of stop band rejection – would require intervening make up gain of up to 40dB, in addition to the typical end to end gain after the mixer of up to 25dB. In direct conversion, however, the digital signal processing required to maintain image rejection in a less than perfect analogue world would require so much number crunching that the low power advantage of direct conversion seems in doubt. Pushing higher At more than 25MHz and 70dB SFDR, active filters become impractical due to gain bandwidth product (GBWP) limitations in amplifiers. There are active filters with 15MHz to 20MHz of usable bandwidth, like Linear Technology's LTC6605 family, but if gain is also required, GBWP requirements are pushed higher. It is often overlooked that the higher the required SFDR, the more demanding the GBWP requirement. And the higher the bandwidth relative to GBWP, the more sensitive gain/phase matching becomes to the GBWP variation of the amplifiers in the active filters. At more than 25MHz, higher order LC filters becoming practical as inductors are reduced to a reasonable size. However, this brings the problem of open circuit magnetics, comparatively lax component tolerances and the potential for unpredictable coupling. If a channel sees coupling from an adjacent channel as one component of I and Q, the modified frequency response due to mutual inductance will modify image rejection, at least toward the upper end of the passband where the frequency response of that channel will be most affected. Another consideration is the common mode produced at the output of the differential filter if one side of the filter experiences coupling from the adjacent filter. This may degrade signal balance to the extent that the common mode component may only be 20dB less than the differential component; enough to compromise channel to channel isolation in a multichannel a/d converter and SFDR. If multichannel a/d converters are not driven with good amplitude and phase balance, there is the risk of inducing ground bounce, which may phase modulate the clock and affect other channels. Pushing the direct conversion architecture to the full Nyquist bandwidth is understandable, but comes with numerous challenges. Many of these could be avoided simply by using a higher sampling rate for the target bandwidth. Derek Redmayne, pictured, is a mixed signal product applications engineer, Alison Steer is a mixed signal product marketing manager. Both are with Linear Technology.