Chuong 9.html

ADSL, VDSL, and Multicarrier Modulation .John A.C.Bingham Copyright # 2000 John Wiley & Sons, Inc. Print ISBN 0-471-29099-8 Electronic ISBN 0-471-20072-7

9

COEXISTENCE OF ADSL WITH OTHER SERVICES

The “other services” to be considered here are voiceband and the three types of BRI (two echo-canceled, one TDD) de®ned in G.961. Coexistence in the ®rst three cases (Sections 9.1 and 9.2) means on the same UTP, with special problems associated with the ®ltering needed to separate the services.In the fourth case (Section 9.3) it means only in the same binder group; the problem is the now familiar NEXT but at such a level that it requires special techniques to make ADSL viable.

9.1 COEXISTENCE WITH VOICE-BAND SERVICES

These have been traditionally called POTS (plain old telephone services), and it has been said that they use the “voice” band.Many of the services are no longer voice, and some of them (e.g., V.90 modems) are neither plain nor old,¹ but nevertheless, we will use the old terminology.Originally, the band extended only to 3.4 kHz, but now with the advent of high-speed modems (paricularly V.90) it extends up to 4 kHz; it would be useful to have the ADSL band extend down as

^{close to the 4 kHz as possible, but 25 to 30 kHz is a practical lower limit. 2}ADSL and POTS are separated by a bidirectional low-pass/high-pass ®lter pair as shown in Figure 9.1. The input impedance of each ®lter must go high out of band (so as not to load the other ®lter), so they must begin with a series inductor/ capacitor as shown. T1.413 and G.992 consider the high-pass ®lter to be part of the ATU, and the term splitterless that is used to describe ADSL lite really means “low-pass ®lterless.” It is more conducive to an understanding of the overall problem, however, to consider the three-port network as one separate device.

If the POTS device is a modem that presents a linear, resistive impedance to the line, the pair needs to perform just three tasks:

¹ [Starr et al., 1999] said that P(ositiveley)A(mazing)N(ew)S(ervices) had been suggested as a name, but I think Tom made that up!
² It is sometimes thought that these ®lters must be passive, but active low-pass ®lters as described in Section 9.1.4 are essential in some telephone systems.

181
Figure 9.1 POTS splitter as a three-port network with low-pass andhigh-pass ®lters.

1.The low-pass protects the ADSL signal from the high-frequency transients (particularly off-hook and ring trip) associated with POTS signaling.
2.The high-pass isolates the voice-band device from the relatively low input impedance (typically, 100 ) of the ATU.
3.The high-pass, together with any internal ®lters, protects the voice-band device from the out-of-band ADSL signals and the ADSL signal from the low-frequency steady-state POTS signaling (dialing and ringing).

If, however, the voice-band device is a telephone, the pair may have two more tasks:

4.The low-pass isolates the ATU from the capacitive input impedance of some telephone handsets, which would otherwise provide a low shunt impedance to the high-frequency ADSL signals.

5.The low-pass isolates the nonlinear input impedance of some telephones from the high-frequency ADSL signals, which might otherwise be intermodulated down into the voice band.

Furthermore, the low-pass must pass the loop current needed for all POTS operations,³and should, ideally, not degrade the voice-band performance.Task 1 for the low-pass is the most dif®cult; performing it would almost certainly ensure performing tasks 4 and 5.For ADSL lite, however, task 1 is greatly relaxed; then task 5 becomes important also.

³ Traditionally speci®ed as a maximum of 100 mA but that can occur only on a short loop, on which the voice quality would otherwise be high.

Location of the Splitter. A “full” splitter that performs all of the foregoing tasks is, as we shall see, a fairly complex device.It was originally intended that it should be placed at the service entrance [a.k.a. network termination (NT) and network interface device (NID)], and that separate and separated⁴ in-house wiring should be used for POTS and ADSL services.This was deemed, albeit grudgingly, to be acceptable for a sophisticated, very high-speed ADSL service (up to 6 ‡ Mbits/s) for which all equipment was to be owned and maintained by the LEC, but it was later realized that it is unacceptable for a more plebian service (ADSL lite) for which the ATU-R is owned and installed by the customer.The simpli®cation and relocation of the low-pass ®lter are discussed in Section 9.1.6.

9.1.1 Transient Protection for the ATU

The worst transient is ring trip, seen at the CO.When the telephone recognizes the ring signal it goes off-hook and draws loop current; this is detected by the circuitry at the CO, which then switches from the ringing signal⁵ to the À48-V battery, generating a quasi step function.If the telephone does not have special circuitry to trip the ring at a zero crossing, it may occasionally trip it near a positive peak, and the transient at the CO will be a nearly instantaneous transition of almost 150 V: current-limited in some circuits, but not in all.

Much of the expertise that was brought to the discussions of this transient during the early work on T1.413 has been lost, and only the conclusion has survived.This was that in order to prevent ADSL errors, the low-pass ®lter should provide between 70 and 80 dB of attenuation at 30 kHz.In Issue 2 of T1.413, with a better understanding of the operation of the FEC function and with more appreciation of the dif®culty of designing and building the ®lters, this requirement was relaxed to 65 dB [including the 8 dB or so of attenuation that requirement was relaxed to 65 dB [including the 8 dB or so of attenuation that requirement was relaxed to 65 dB [including the 8 dB or so of attenuation that (at 30 kHz) loop].Ring trip seen at the RT is attenuated somewhat, but the 65-dB attenuation requirement was maintained.

NOTE: This requirement for 65 dB at RT assumes that the downstream band extends down to 30 kHz (i.e., EC is used). At about 150 kHz, which is the low end of the downstream band for an FDD system, all POTS transients would be further attenuated, and furthermore, the low-pass ®lter could be designed to have considerably more attenuation.⁶ I have not seen a realistic assessment of the low-pass attenuation (different at CO and RT) needed for an FDD system.

⁴ Physical separation of the pairs is necessary to prevent the POTS signaling transients from NEXTing into the ATU-R.
⁵ 20 Hz in the United States, only slightly (insigni®cantly for our present purposes) different elsewhere in the world.
⁶ The second-order mini®lters discussed in Section 9.1.6 would have 28 dB more.

9.1.2 Isolating the Voice Band from the (Low) Input Impedance of the ATU

This is a very dif®cult requirement to quantify.As noted above, T1.413 and G.992 consider the high-pass ®lter to be part of the ATU and the low-pass ®lter to be a separate device that must be speci®ed independently.Therefore, they de®ne the input impedance of the high-pass/ATU that must be assumed when designing the low-pass ®lter.This was probably the only practical way to deal with the problem, but the purpose of the isolation is to help maintain the quality of the voice-band service, and this can be done better if the two ®lters are designed together. The problems are considered all together in Section 9.1.5.

9.1.3 Maintaining Voice-Band Quality

The three characteristics of voice-band performance that should, ideally, be maintained are noise level, end-to-end response, and return loss.If the loop was loaded before the ADSL service was installed, the question for response and return loss is: What should be maintained? The rules for adding loading coilsÐ and therefore the quality of the “pre-ADSL” service to which the customer is accustomedÐseem to vary from LEC to LEC: Some say any loop over 12 kft may be loaded; others say only loops over 18 kft are loaded.

Noise Level. The attenuation of the lower sidelobes of the upstream signal and the high-pass ®lter should together, ideally, keep all ADSL signals out of the voice band.If, however, the input impedance of the telephone is nonlinear, components above 30 kHz can be intermodulated into the voice band; task 4 in Section 9.1 becomes important. If there is no low-pass ®lter, the upstream transmit level must be reduced to less than À46 dBm/Hz (a reduction of at least 8 dB) to keep the intermodulation products to a tolerable level.

End-to-End Response. Figure 9.2 shows the response of 13.5 kft of 26 AWG (one of the test loops de®ned in T1.413) both loaded and unloaded. It is clear that removing the loading coils greatly degrades the response.It is impracticalÐ perhaps even theoretically impossibleÐto restore the response to that with loading coils, but it is possible to reduce the roll-off slightly by using mini®lters, as shown by the third plot (marked with asterisks) and discussed in Section 9.1.6.

Return Loss. In the United States, three types of return losses are de®ned for voice use:

1. Singing return loss (SRL) low: the minimum return loss in the band 300 to 500 Hz
2. Echo return loss (ERL): the “average” return loss (obtained by comparing total transmitted and re¯ected powers) in the band 500 to 2500 Hz
3. SRL high: the minimum return loss in the band 2500 to 3400 Hz

Figure 9.2 Voice-band responses of 13.5 kft of 26 AWG: loaded and unloaded.
ITU Recommendations G.122 and G.131,⁷ however, de®ne the band for SRL high all the way out to 4.0 kHz.
The problem of controlling these is greatly complicated in the United States by two factors:

1.Even though, after the loading coils have been removed, a loop is almost symmetrical in the voice band,⁸ different terminating impedances are used at the CO and RT: 900 in series with 2.16 mF, and 600 , respectively. This makes the input impedances seen at the two ends different and requires different ®lters at CO and RT.

2.The terminating impedances are resistive (or almost so) instead of a ®rstorder bilinear RC (RRC) approximation to the characteristic impedance that is used in the United Kingdom and many other European countries. This means that the input impedances vary considerably with loop length, and all designs have to be compromises.

At both ends of the loop, attempts are made to match the loop impedance and balance the 4W/2W network.In a telephone this reduces the sidetone from
⁷ This is a very old reference, which I got from [Freeman, 1981]; it may be out of date. ⁸ Gauge changes and bridge taps have very little effect in the voice band.
Figure 9.3 Compromise RRC impedances: (a) U.S. CO and RT; (b) U.K.

microphone to speaker; in a modem it lightens the load of the echo canceler, andÐprobably most importantÐat the CO it reduces the delayed echo that the far end experiences.The compromise matching impedances are RRC as shown in Figure 9.3.

The return losses at the CO relative to the RRC impedance of Figure 9.3( a) for the loaded and unloaded 13.5-kft loops are shown in Table 9.1, together with the requirements in Issue 1 of T1.413 with the splitter installed.It can be seen that the loaded loop does not completely meet the requirements even without the ADSL service.

Figure 9.4 may help to explain what is happening; it shows the loci with frequency of the RRC impedance and the input impedance at the CO of both loops.It can be seen that the impedance with loading coils is very different from the RRC impedance.The return loss situation is thus the reverse of the response one: it is poor with loading coils and good when they are removed.

The situation deteriorates again, however, when LC low-pass ®lters are inserted.The requirement of 65 dB attenuation at 30 kHz can be met with a fourth-order elliptic function ®lter (C041014 in [Zverev, 1967]) with a cutoff fourth-order elliptic function ®lter (C041014 in [Zverev, 1967]) with a cutoff resistance the locus of the input impedance would be a small quasi-ellipse, and the minimum return loss would be 20 dB.If this ®lter is terminated by the loop, however, then, as pointed out in

TABLE 9.1 Voice-Band Return Losses for Loaded and Unloaded Loops SRL Low ERL SRL to 3.4 kHz SRL to 4.0 kHz (dB) (dB) (dB) (dB)

13.5 kft 26 AWG unloaded 8.9 12.1 16.4 16.4
13.5 kft 26 AWG loaded 8.9 10.7 2.8 2.8
T1.413 requirements 5 8 5 n/a

Figure 9.4 Input impedance at CO of 13.5-kft loops with impedance of compromise RRC.

[Cook, 1994], the locus expands greatly, and the return loss relative to the compromise RRC impedance falls to nearly 0 dB at the edge of the passband. Figure 9.5 shows the input impedances of a 9-kft 26-AWG loop (short enough never to have needed loading) with no ®lter and the above fourth-order ®lter. Table 9.1 also shows the return losses for this loop: with and without ®lters.

As a way of explaining this effect, Cook suggested that the ®lter behaves in its passband like a lossless transmission line.At dc and all other frequencies at which l ˆ n, the ®lter is transparent and Z_in ˆ Z_loop, but at frequencies at which lˆ…n ‡ 1=2† the ®lter has the well-known effect of inverting its terminating impedance (i.e., Z_in ˆ Z ²=Z_L, and for example, an open circuit₀
looks like a short circuit).This means that at the latter frequencies the negative imaginary part of the “RC” impedance is transformed to positive.Thus the input impedance is RLC and oscillates between these two extremes; the number of the oscillations and the frequencies at which they occur depend on the order of the ®lter, the cutoff frequency, and the sharpness of the cutoff.Some improvement in the return loss can be achieved by increasing the order of the ®lter⁹ and pushing the cutoff frequency much higher, but my own calculations and published results

⁹ A sixth-order quasi-antimetric ®lter is common in the Unites States. Figure 9.5 ZinRT of 9-kft 26 AWG with andwithout fourth-order ®lter plus compromise RRC.
[Cook, 1994] and [Hohhof, 1994] suggest that there is no standard “textbook” ®lter that meets all requirements.
9.1.4 One Solution to the Impedance Problem: Generalized Immittance Converters

The most elegant solution to the impedance problem as it exists in the United Kingdom was described in [Cook, 1994] and [Cook and Sheppard, 1995]. Loading coils are not used in the United Kingdom, so all loop input impedances are RC, and both the line-driving/terminating impedance and the balancing impedance in the 4W/2W network are compromise RRC, such as are shown in Figure 9.3(b): trying to match the characteristic impedance of the loop.Many of the telephones are active and require a return loss of about 18 dB or better.

A generalized immittance converter (GIC) is a reciprocal active two-port de®ned by
1 Z_in1 ˆ F₁… p†Z_load2 and Z_in2 ˆ_{F1… p†}Z_load1 …9:1†
Figure 9.6 Impedance transformations using GICs.
and the solution uses a mirror-image pair of them at each end arranged as shown in Figure 9.6. The requirements for both GICs are
Z^{RI… p ‡ p1†} and Z_in2 ˆ R_LC …9:2†_in1 ˆ_{p ‡ p2}

In-line GIC. One implementation uses the two-ampli®er circuit described in [Antoniou, 1969] and [Fliege, 1973].¹⁰ This is shown in its most basic form in Figure 9.7. The input, output, and “bc” nodes are connected to the inputs of the two op-amps so as to keep the three nodal voltages V₁, V₂, and V_bcequal.Then the impedances seen at the two ports are

Z_in1 ˆ^ZaZc Z_L2 and reciprocally Z_in2 ˆ^ZbZdZ_L1 …9:3†_{ZbZd ZaZc}

If the ampli®ers are considered to be ideal, there are (theoretically) many ways of connecting the nodes to the op-amp inputs and assigning the pole and zero of (9.2) to Z_{1;2;3; and 4}, but the requirement of high-frequency stability with real

Figure 9.7 Two-ampli®er implementation of a GIC. ¹⁰ Variations of this circuit are also used to implement biquadratic ®lter sections, gyrators, and frequency-dependent negative resistors.
ampli®ers reduces the number of workable combinations to a very few; unfortunately, details of the preferred circuits are proprietary and unpublished.

Off-line GIC. The Fliege circuit maintains V₁ ˆ V₂ and scales the currents; the circuit described in [Cook, 1994] on the other hand, maintains i_i ˆ i₂ and scales the voltages.It does this by adding a voltage in series between the ports:

V
2
ˆ
V
1
‡
V
R_I… p ‡ p₁^†
1_{RLC… p ‡ p2}†À ^{1 …9:4†}
It is desirable that the added part in brackets be zero at in®nite frequency (i.e., that the active circuitry be low-pass).Therefore, R_LC is set ˆ R_I, and
V^{p1À p2} …9:5†₂ ˆ V₁‡ V_{1 p ‡ p2}

Implementation. V₁ can be detected from the balanced port 1 by a highimpedance (so as not to load the port) difference circuit, multiplied by the ®rstorder low-pass transfer function in (9.5), and then added in to the series voltage via a third winding on a balanced (i.e., with zero differential-mode inductance) “summing” inductor.This is shown in Figure 9.8, which is a simpli®ed version of Figure 8 of [Cook, 1994].

Both low- and high-pass ®lters must be balanced about ground.Figure 9.8 shows how each series inductor can be economically implemented as a balanced pair wound on the same core.The shunt capacitors can be implemented as one (C₁) between the balanced arms or as two double-size capacitors (2C₂ and 2C₂) to “ground”.The relative merits of these two depends on the need for commonmode ®ltering and on the availability of a good ground; analog designers do not all agree on which is better.

The advantage of the “off-line” circuit is its robustness; if power is lost, the output impedance of the driver op-amps goes high, the third winding on the summing inductor is open circuited, and the GIC has no effect; the voice-band signal is maintained, albeit with degraded impedances.

Figure 9.8 Balancedimplementation of low-pass and“off-line” GIC.

Use of the GIC Approach in the United States. It was suggested in [Cook, 1994] that U.S. needs (with quasi-resistive terminations) could be met with just one GIC on the line side of the ®lter at each end.This, however, would only partially solve the problem: the impedance seen from the telephone or the CO would now be resistive instead of the expected RRC.The return loss of this relative to the RRC compromise would be better than it would be with no GIC at all, but whether it would be good enough requires careful study.

9.1.5 A Partial Solution: Custom Design by Optimization

As noted in Section 9.1.3, it seems fairly certain that no pair of conventional (designed to work between resistive impedances) LC ®lters can meet both the attenuation and the return loss requirements.However, using a pair of these ®lters as a starting point, and iterating on their components simultaneously to minimize a weighted sum of errors (low-pass passband response and return loss and stopband response, and high-pass response), can result in a signi®cant improvement.

NOTE: The weightings will almost certainly have to be adjusted as the iteration progresses, so as to keep each of the performance parameters reasonably within bounds.

UnfortunatelyÐfor the bookÐprograms to do this iterative design are proprietary.All I can say (as an incentive!) is that echo return losses of 12 dB can be obtained with most loops using a sixth-order low-pass.This does, however, require simultaneous adaptation of both low-pass and high-pass; it is doubtful that 12 dB can be achieved when the low-pass is designed as a separate unit and a ®xed compromise input impedance is used to emulate the high-pass.

The resulting passband responses are far from equal ripple: Loss and mismatch at low frequencies seem to push out the frequencies at which lˆ…n ‡ ¹† and to “postpone” the input impedance becoming inductive.The₂
return losses are not as good as could be obtained with an unloaded loop and no ®lters, but they are considerably better than could be obtained (in the United States at least) with loaded loops.The passband response is not as good as with a loaded loop, but it is probably better than with an unloaded loop without a ®lter.

9.1.6 Simpli®ed (Dispersed and Proliferated) Low-Pass Filters¹¹

One way of avoiding the dreaded “truck roll” would be for a customer to buy (full-size) low-pass ®lters and install them at every POTS device; the transients would be ®ltered at the source, and all POTS-associated signals on the house wiring would be con®ned to the voice band.These ®lters, however, have a minimum of three inductors each and are bulky and expensive.

¹¹ These have been called in-line ®lters, but this is a misleading name; they are not in the line (loop?), they are associated with every POTS device. Mini®lters is a better name.

Early expectations for ADSL lite were that these ®lters could be eliminated altogether, and the only problem would be the short bursts of errors (perhaps correctable, but if not, then probably tolerable) caused by POTS signaling. It was soon realized, however, that tasks 4 and 5 in Section 9.1 are ongoing; with no low-pass ®lters some telephones would receive an intolerable level of noise whenever the ATU-R was transmitting,¹² and some would almost short out the ADSL loop at high frequencies whenever they were off hook.

NOTE: The difference between protecting against transients (task 1) and against steady-state effects (tasks 4 and 5) is important for the TC layer.For many applications a burst of errors upon ring trip would be tolerable (as long as neither layer, PMD or TC, lost “sync”), but a subsequent change of state that resulted in a signi®cantly lower capacity would be very dif®cult to deal with.

An attractive compromise between full-size ®lter and no ®lter is a simple second-order (one inductor, one capacitor) mini®lter that performs tasks 4 and 5
and probably protects against most signaling transients from most phones.¹³
Figure 9.9 shows a typical house wiring with mini®lters plugged into every RJ11
jack.

Figure 9.9 Typical in-premises wiring with mini low-pass ®lters.

¹² There was a suggestion to reduce the level of upstream transmission, but this would reduce the range when the upstream was subjected to alien NEXT.
¹³ “Probably” and “most” are very imprecise words, but they are the best I can do at this time. Tests are needed with many different loops and phones.

NOTE: These ®lters must be balanced, and because there will not usually be a ground available at the jack, the shunt capacitor must be a single one between the balanced arms; no common-mode ®ltering can be performed.

A typical ®lter would have a nominal cutoff frequency of about 4 kHz (achieving about 34 dB of transient protection at 30 kHz), with L % 28 mH and C % 50 nF.The magnitude of the input impedance from the loop side with a telephone-induced shunt capacitor load %5k at 30 kHz, so the load on the at 30 kHz, so the load on the loop is negligible (task 4).

Input Impedance and Return Loss. Even such a minimal ®lter still rotates the RRC loop impedance slightly, and the return losses above about 2 kHz are between those for full-size and no ®lters.This is shown in Figure 9.10 for a 9-kft 26 AWG loop, and the return losses are shown in Table 9.2

Response. A bonus feature of mini®lters is that they improve the voice-band response slightly.The series inductors act somewhat like loading coils and partially cancel the shunt capacitance of the loop.The third curve in Figure 9.2 shows the response of the 13.5-kft 26 AWG loop with two mini®lters added. This loop may have been loaded before ADSL installation, and the customer may

Figure 9.10 ZinRT of 9-kft 26 AWG with mini®lter plus compromise RRC. TABLE 9.2 Voice-Band Return Losses with and Without Filters

SRL Low (dB)
ERL SRL to 3.4 kHz SRL to 4.0 kHz (dB) (dB) (dB)

9 kft 26 AWG
Without ®lter 19.1
With full ®lter 17.7
With mini®lter 19.7
18.6 17.6 17.3 10.6 2.4 1.5 18.4 9.6 6.9

have become accustomed to a ¯at response, so even a slight ¯attening may be welcome.

Mini®lters at the CO. All the emphasis has previously been on simplifying the remote ®lters, and the tacit assumption seems to have been that the CO would still use a full-sized ®lter.In view of the better return loss achieved with the mini®lters, a strong argument could be made for using them at both ends.The high-frequency components of the signaling transients imposed on the upstream signal would, of course, be increased; whether they are tolerable may depend on whether interleaving is used for upstream (see Section 2.4.3).

Mini®lters with Active Telephones and/or Complex Terminating Impedances. Table 9.3 shows the four return losses (at either CO or RT because the complete circuit is now symmetrical) for 9-kft and 13.5-kft 26-AWG loops¹⁴ with terminating and reference impedances as shown in Figure 9.4(b): with no ®lter, a second-order ®lter, and a fourth-order ®lter.It can be seen that for the average-length loop the mini®lter actually improves the return loss over most of the passband; for the long loop it improves the ERL but signi®cantly worsens the SRL high.It remains to be seen whether the SRL high can be either improved or

TABLE 9.3 Return Losses with Complex Terminations: with and Without Filters SRL Low ERL SRL to 3.4 kHz SRL to 4.0 kHz (dB) (dB) (dB) (dB)

9 kft 26 AWG Without ®lter With full ®lter With mini-®lter

13.5 kft 26 AWG Without ®lter With full ®lter With mini®lter 15.2 (note) 18.6 15.1 14.4 15.1 15.5 3.2 1.7 15.7 22.5 18.0 8.9

13.0 17.9 15.3 14.7
13.1 12.5 2.9 2.6
13.4 21.9 9.1 5.9

Note: The values without any ®lters are probably slightly too low because the compromise impedance is for UK cable, andthe RLGC parameters usedwere for U.S. cable!

¹⁴ I know! Loops in the United Kingdom are measured in kilometers and they are metric gauge, but since the point here is comparison with what we have discussed previously, it is more informative to stay with U.S. units.

tolerated, and whether mini®lters can be used instead of full-size ®lters and GICs.
9.2 G.992 ANNEX B: COEXISTENCE WITH ECHO-CANCELED ISDN

In many countries with a large installed base of ISDN modems, it is necessary that ADSL operate in the frequency band above ISDN.The transmit signals as de®ned in T1.601 and G.961 Annexes I and II (2B1Q and 4B3T) have nominal À3-dB points at 80 and 120 kHz, respectively, and are only very lightly ®ltered, so if the crossover frequency between ISDN and ADSL had been continuously variable, it would have been hard to get agreement on the permissible low end of the ADSL.DMT as de®ned in T1.413 had, however, already established bands of 138 kHz (comprising 32 subcarriers): band 1 for upstream, bands 2 through 8 for FDDed downstream.It was very easy (and probably nearly optimal) to reserve the band 1 for ISDN, band 2 for upstream, and bands 3 through 8 for downstream.¹⁵

The most straightforward way of encoding the upstream data into band 2 is to constellation-encode just as for a T1.413 or G.992.1 Annex I modem, and then modulate onto tones 33 through 63 using a 128-pt IFFT with tones 1 through 32 zeroed.However, before the issue of ADSL over ISDN was raised, at least one manufacturer had ®rmly embedded a 64-pt FFT in silicon, and their only feasible way of moving the data up to band 2 (and down again in the receiver via the reverse process) was to modulate it to the lower sideband of an IF carrier at 276 kHz (very easily generated).This has the effects of interchanging the sidebands of each narrow QAM signal and thereby changing the constellation encoding rules.ETSI TM 6 and ITU SG 15 are currently working on the problem of ensuring compatibility between a transmitter that uses one method and a receiver that uses the other.

9.3 G.992 ANNEX C: COEXISTENCE WITH TDD ISDN

TDD¹⁶ ISDN as de®ned in Annex III of G.961,which is used primarily in Japan, uses a bipolar [a.k.a. alternate mark inversion or (AMI)] pulse at a symbol rate of 320 kbaud with very little ®ltering.The transmit PSD is shown in Figure 9.11 together with those of HDSL and downstream ADSL for comparison.The total data rate, which is slightly less than 320 kbit/s because of a guard period to allow for propagation delay, is shared equally between upstream and downstream using TDD with a superframe period of 2.5 ms (rate ˆ 400 Hz).The XT coupling

¹⁵ Kindred NEXT would make the band above 138 kHz unusable for full-duplex transmission, so there can be no EC option for ADSL above ISDN.
¹⁶ This was originally called time compression multiplex (TCM), but for ADSL TCM means trellis coded modulation.The ITU has, unfortunately, perpetuated the use of the confusing TCM, but I will stick with TDD.

Figure 9.11 PSDs of TDD ISDN, HDSL, anddownstream ADSL.
coef®cients are given in Table 3.6; they are higher than anything encountered previously.

ADSL service in the same cable as TDD ISDN is very differentÐboth in crosstalk conditions and in the duplexing method that those conditions dictateÐ from any considered previously.The main problem is the time-varying nature of the crosstalk; when the downstream ADSL is going “with the TDD ¯ow” it incurs alien FEXT, which is milder than kindred FEXT, but when it is going against the ¯ow it incurs alien NEXT, which may be very severe.Therefore, ADSL may use at least two different bit loadings: heavy for FEXT and light for NEXT.If the ADSL and TDD clocks are not frequency-locked, the ADSL symbols will precess slowly through a TDD superframe and incur changing crosstalk.A continually changing bit loading to match this would be very inconvenient, so either the clocks must be locked or the light loading must be used whenever there is any NEXT (protecting the error rate, but inef®cient).

9.3.1 Synchronizing TDD ISDN and ADSL

Synchronization is very easyÐthe 400-Hz clock is available at the CO for all ATU-Cs to useÐbut a crucial question is: What should be synchronized? The data symbol rate for ADSL is 4 kBaud, exactly 10 times the frame rate for TDD ISDN. The on-line symbol rate as de®ned by T1.413 and G.992 Annex A, however, is (69/68) Â 4 kBaud, to allow for the insertion of one sync symbol after every 68 data symbols.Therefore, if the symbol length (544 samples) of Annex A is preserved, the on-line symbols will precess slowly with respect to the TDD ISDN, and line up only once every 690 symbols: the synchronization must be 10 ADSL superframes (690 symbols) ˆ 69 TDD frames.

Unfortunately, the 544-sample symbol was preserved¹⁷ in Annex C, and the result is a complicated and inef®cient framing structure.It is complicated because it must keep track of the crosstalk/loading status of each of the 690 symbols of a TDD hyperframe, and it is inef®cient because approximately 20% of the ADSL symbols are “transitional”; that is, they incur some FEXT and some NEXT, and because of the binary (heavy/light) loading scheme, they must be classi®ed as light.¹⁸ The reader is referred to G.992 for all the details.¹⁹

How It Should Have Been (Could Still Be?). The sync symbol was devised in 1992 and included in the DMT standard just in case it was needed to deal with loss of symbol synchronization.The original cyclic pre®x of 40 samples was shortened to 32 [see equation (8.5b)], and the saved samples accumulated for 68 symbols to form the sync symbol.Subsequently, as synchronization methods were developed, it was realized that the sync symbol was not really needed.At different times during the discussions on T1.413 Issue 2 there were proposals to remove it, but it was decided that the 1/69 gain in capacity was not worth the inconvenience of backward incompatibility.

When the problem of synchronizing with TDD ISDN’s 400 Hz arose, however, there was no need for a strict adherence to T1.413 with its cyclic pre®x of 32 samples.²⁰ The simplest solution would have been to discard the sync symbol and revert to a cyclic pre®x of 40 samples²¹ (see Table 8.1) with the data symbol rate ˆ the on-line symbol rate ˆ 4 kBaud; there would then be exactly 10 ADSL symbols in one TDD frame.The crosstalk/loading status would repeat every 10 symbols, and there would be no need for any transitional symbols.

A possible timing diagram is shown in Figure 9.12, where, as in G.961, time is measured in unit intervals (UIs) of 3.125 ms, and the propagation on the longest loop (9 km) is assumed to be 50 ms (16 UIs).It can be seen that if the ATU-C transmitter begins its superframe 16 UIs before the ISDN transmitter, the downstream superframe can comprise ®ve heavily loaded and ®ve lightly loaded symbols, and the upstream superframe, four and six, respectively.

¹⁷ Probably an in¯uential company had a 32-sample cyclic pre®x cast in silicon and was too shortsighted to abandon it: an egregious example of commercial muscle defeating technical merit!
¹⁸ According to Annex C the system is even less ef®cient than that; out of 345 symbols only 124 are allotted for heavy loading (when there is only FEXT) and 216 for light (when there is any NEXT).I do not understand this.
¹⁹ You can tell my heart is not in it!
²⁰ There would be no problems of end-to-end compatibility here; if one ATU has to coexist with TDD ISDN, so does the other.
²¹ With a small reduction in distortion as a spin-off bene®t.

Figure 9.12 Ideal timing diagram for one TDD ISDN superframe, and ADSL synchronized to it.
9.3.2 Band Assignments and FFT Sizes

Because the crosstalk is the same for downstream and upstream, the channel capacity is also.The original proposal was to use the full bandwidth for both directions in a TDD mode, but because (1) the upstream rate need be only oneeighth of the downstream, and (2) Annex A of G.992 de®nes a conventional FDD system with band 1 allotted to upstream and bands 2 through 8 allotted to downstream, it was decided to stay with this band assignment and use dual bit loadings in both directions.

How It Might Have Been. Annex A FDD with dual bit loading is inef®cient because

^* During ISDN up the downstream capacity in band 2 (light loading) ( the upstream capacity (potential heavy loading)
^* Conversely, during ISDN down the upstream capacity in band 1 (light loading) ( the downstream capacity (potential heavy loading).

TABLE 9.4 Recommended Band Assignments for ADSL with TDD ISDN Crosstalk During Mandatory Optional

ISDN up Upstream in band1 Upstream in band2 (note 1) Perhaps

Downstream in bands 3±8 (light loading) (note 2)

ISDN down Downstream in bands 1±8

Notes: 1. Annex A requires only a 64-pt (I)FFT, andAnnex B could be implementedthat way, but for all the reasons discussed in Section 8.3.4, a 128-pt (I)FFT will probably become the norm.
2. Some existing upstream ISDN receivers may have very little receive ®ltering andtherefore not be able to tolerate the high levels of out-of-bandNEXT causedby ADSL downstream transmission against the ¯ow.

Therefore, considerably higher data rates could be achieved by the assignments shown in Table 9.4. Thus dual bit loading would never be needed for upstream. For downstream it should be an option²² (controllable by system management); its advantages (A) and disadvantages (D) are as follows:

D.If quad separation is not implemented, the small increase in capacity achieved by transmission against the ¯ow in bands 3 through 8 may not be worth the additions to the training sequence that would be needed to calculate and transmit the extra bit table.

A.If the TC and higher layers are able to take advantage of periods of higher data rates, dual bit loading should be retained for those times when there is very little or no active ISDN in the cable: the light becomes much heavier.

D.Without dual bit loading in either direction the system becomes pure TDD, which could be implemented cheaply with a single (I)FFT processor in each ATU.

9.3.3 Separate Quads for ISDN and ADSL

If an overall DSL system could be managed so as to put ISDN and ADSL in separate quads, severe same-quad alien crosstalk could be avoided, and the worst-case alien XT coef®cents would be lower than for unquadded cable; dual bit loading for downstream would be more advantageous.

9.3.4 ULFEXT from Close-in ISDN Modems

Usually, the upstream ADSL capacity will be greater than one-eighth of the downstream because the upstream uses two bands instead of the usual one.If, however, some of the ISDN loops are short, there is a potential for high levels of

²² The training sequence should be arranged so that if one of the ATUs does not have the capability, the other should use a zero second loading.

FEXT into upstream ADSL because TDD ISDN does not use any power cut back.The danger of this effect would seem to be a strong argument for separate quads, but separation may occur naturally without any conscious management: loops in the same quad would probably be of similar lengths.