## Friday, March 29, 2013

### You can't touch this, Part II

In my previous post, I reviewed the concept of Average and Normalized Power, more as an introduction to some further thoughts about the topic of NP Busters. I also said that this would be a two part discussion, with Part II on the topic of NP Busters. Well I am getting to that but it will actually require three parts, so here continues the discussion on Normalized Power, as another prequel to an NP Buster chat. I will at least introduce what is meant by an NP Buster.

Previously I demonstrated by way of an example of a proposed interval session how average power can be a misleading indicator of metabolic strain, especially when power output is highly variable, and that Normalized Power represents a better means of measuring metabolic strain. Well we don't need to make up theoretical examples, we can turn to real data.

### Criteriums versus Time Trials

Let's consider the Normalized and Average power from hard rides of different types but of similar durations. An obvious example would be to compare a time trial with a criterium race.

A TT is typically ridden solo and involves sustaining a high power in a relatively steady state manner, with perhaps some variability if the terrain is not flat or has some technical elements, while a criterium involves substantially variable power outputs as one deals with or dishes out the attacks and surges, the braking and/or coasting into and accelerations out of turns, the inevitable driving of the pace in or to establish a break, and sitting in the slipstream of others when recovering. As rides, they are poles apart.

The following chart (click on it to see a larger version) shows a comparison of the power output over time for a time trial and a criterium race by the same rider, performed within about five weeks of each other and both on relatively flat courses. There are two plots for each race. The lines that jumps up and down are the second by second power data trace, and the two straight horizontal lines are the average power from each race. The time trial (blue) is a little shorter in duration than the criterium (red).

The instantaneous power output is a little hard to follow since it jumps up and down so much, but even so, it's clear that the criterium power line (red) is far more variable than the time trial power line (blue). This is pretty typical. So while both of these races were hard efforts by the same rider and over reasonably similar durations, there was a substantial 40 watt difference in the average power.

On closer inspection we can see a period in the crit race from around the 33-minute mark where power dropped substantially. It happened that the rider had a puncture and "took a lap out" to replace a wheel and rejoin the race (annoyingly as they had established a breakaway prior to that). So we would expect this lower power period would account for some of the lower average power overall, even so, the average power up to that point was 272W, still 25W less than the average power in the time trial.

But let's not forget that time spent not pedalling affects what you can do when you are pedalling, and so that mini break no doubt meant a little freshening up before rejoining the race, and an ability to go a little harder than might have been the case with no recovery.

A good way to gain some insight is to view the power trace after applying a filter to the data, and one simple filter is a rolling 30-second average (i.e. each point on the chart represents the average power for the preceding 30-seconds). Here's the same plot showing the rolling 30-second average power:

The vertical scale is now halved which means variances are amplified. The 30-second rolling average makes it easy to spot differences in the power sustained during sections of a ride. In this example we can readily identify periods during the criterium of sustained harder and easier effort. Likewise, the time trial also shows two brief drops in power output, which correspond to a steep decline on the course with speeds too fast for continued pedalling.

A 30-second rolling average power filter is of particular interest as metabolic responses to changes in effort really start to kick in at around that time frame - many have what we call a "half-life" of around 30-60 seconds. Very brief forays (a handful of seconds) at higher powers are not all that metabolically stressful but sustain the higher power for longer (>20-30 seconds) and it gets ugly, fast. How fast depends on how hard you go.

Hence it's no coincidence the algorithm used to calculated Normalized Power is based (partly) on a rolling 30-second average power filter. There's a couple more important elements to the NP formula than that (although it's not a very complicated formula) but it starts with this 30-second rolling average.

So what was the Normalized Power for these two races? Well here they are plotted on the chart as the two horizontal lines:

In effect, the Normalized Power from each race was the same (OK, one watt different). So even though the races were very different in style, they were both hard and produced a Normalized Power that was more representative of the metabolic strain experienced.

OK, so that's pretty nifty, and is why Normalized Power is a good way to glean from races how your fitness is tracking despite the lack of a formal testing protocol.

It should also be of no surprise there is very little difference between the Average and Normalized Power for the time trial (297W and 299W respectively), since the effort was already relatively steady state, and NP is about providing a steady state power equivalent (hence the name "Normalized").

By definition, Normalized Power will be equal to or greater than Average Power, and the gap between them will depend on the amount of variability there is in the rolling 30-second power, and especially the duration and number of forays at very high power levels.

### Using Normalized Power to estimate Functional Threshold Power

Since Normalized Power is providing a steady state power equivalent for longer (dominantly aerobic) durations, then it follows that one can consider NP from hard rides/races of about an hour as one means to estimate FTP.

The well established rule of thumb is for durations of about an hour, Normalized Power will be no more than 5% higher than the maximal quasi-steady state power a rider is truly capable of. Since maximal quasi-steady state power for about an hour is the definition of Functional Threshold Power, then we can simply state:

~1-hour NP <= 105% of FTP

or at least that it will be for the large majority of riders, a large majority of the time.

So if you notice from a hard ride/race of about an hour that NP is > 105% of FTP, then it's quite possible your FTP is higher than you think it is.

There are of course caveats to this rule of thumb. I'll go over these as they impact the definition of an NP Buster and can help explain what some perceive to be anomalies when interpreting their own NP numbers.

The duration caveat
Since we are primarily concerned with obtaining a measure of equivalent aerobic metabolic demand/strain, then the duration of any comparison of highly variable versus steady state efforts needs to be sufficiently long to reduce the confounding impacts from individual differences in anaerobic work capacity and neuromuscular power capabilities relative to a rider's aerobic capabilities.

For this reason, NP numbers from rides or parts of a ride of less than 20-minutes duration are not suitable for such comparisons, nor as an indicator of a metabolic steady state power equivalent. I generally take more notice of NP for durations of at least 30-minutes, but it depends on the rider's individual circumstances and capabilities. As the duration of a ride reduces (e.g. down towards 20-minutes), then the difference between NP and a rider's actual maximal steady state power can become somewhat wider.

The circumstantial caveat
There are circumstances where no matter how one rode (steady state or variable), their power output would be somewhat different when compared to another circumstance. Examples of this might be comparing riding on an indoor trainer to an outdoor ride as some people experience a sizeable difference in the power they can sustain indoors versus out.

Another might be comparing long steep hillclimb to flat terrain, or on a road race bike versus an aggressive time trial bike position that might compromise power output for some aerodynamic gains, or really hot day, or at altitude and so on. Another is the use of frequent out of the saddle efforts engaging upper body musculature versus staying in the saddle.

So while Normalized Power enables a comparison of some apples with some oranges, we need to be thoughtful when using it to compare all types of fruit.

The power meter data accuracy caveat
Well it should go without saying that power data needs to be accurate for the interpretation to make sense. While basic accuracy is a factor, there are ways in which data integrity can be compromised even though the individual data points might still be accurate. This mostly concerns the way some power meter head units collect and store data, especially the sampling rate. If the fruit is bad, well no point in trying to use it.

An example of this is/was Garmin's use of "smart recording", which should in current firmware versions be automatically disabled when using a power meter, but it makes sense to ensure it really is disabled. This was also a factor for older model power meters with memory space restrictions, and options to "down-sample" data (e.g. older Powertap head units). You could get away with 2-second sampling (just), but any more than that would compromise data integrity to the extent that the data might not be all that useful.

The software algorithm caveat
While the Normalized Power algorithm is pretty straightforward and in the public domain, not all software (be it commercial desktop software such as WKO+, home designed spreadsheets or websites) produce the same results. There may be a number of reasons for that, e.g. use of an incorrect algorithm (I've seen it many times with people claiming an NP that was incorrectly calculated) or more subtle matters such as how gaps in power data or variable duration time stamps are handled.

So when doing such analyses and/or comparisons, then consider the software you are using as well and validate it is correctly applying the algorithm. Some food processors take the goodness out of the fruit.

### So what is an NP Buster?

An NP Buster is a ride that breaks the rule of thumb, or put this way:

~1-hour NP > 105% of FTP

provided:
1. the above caveats are taken into consideration (especially power data accuracy, correct calculation of NP, but also the circumstantial caveats), and
2. FTP at around the time of the claimed buster ride has been well established using one or all of Andy Coggan's Sins 5, 6 and 7 referenced in this post on establishing Functional Threshold Power, i.e.:
• using critical power testing and analysis
• from the power that you can routinely generate during long intervals done in training
• from the average power during a ~1-hour TT

Such NP Buster rides have occurred, and there are riders who can produce them. They are however rare, and I'll talk more about them in Part III.

## Wednesday, March 27, 2013

### NP Busters

are the spark for today's musing. It's an old topic but a fun one. I am however going to break this into two parts, first (Part I) to review the concept of Average and Normalized Power, and then (Part II) to chat a little on NP Busters.

An NP Buster?

So before getting into the discussion of NP Busters and just WTF I'm on about, let's just go back to Power 411 to remind us what Average and Normalized Power is all about. This is mostly for those that are new to the concepts, even though NP has been with us for a decade, the number of people beginning to use power in training and racing is ever growing and besides, a refresher is never a bad idea.

For those well versed in power meter analysis and associated software, they are no doubt familiar with the concept of Normalized Power and perhaps don't need to go over old ground the rest of this post covers. Much of what I am covering in Part I is also in this original item by Andy Coggan introducing Normalized Power. I suggest reading it if you have not done so before (and you're interested in learning about this stuff).
In summary, Normalized Power is neat a way of enabling us to make sense of rides that are, by their nature, highly variable in power output, especially when a straight numerical average of a rider's power output is often not that helpful in assessing the "damage" done during a ride.
With that said, you can wait for Part II, the NP Buster chat, or read on...

### Average Power

Average Power is by definition fairly straightforward – being the average of a rider’s moment by moment power output over part or whole of a ride. For example, 5-minutes at 100 watts followed by 5-minutes at 200 watts equates to a 10-minute Average Power of 150 watts.

A measure of work done
Average Power tells us how much mechanical work was performed during a ride. This knowledge has numerous benefits, in particular when assessing daily energy intake requirements:
Average Power (watts) x Ride Duration (seconds) = Mechanical Work Performed (joules).
e.g. 150 watts x 600 seconds (10-minutes) = 90,000 joules (90kJ)

Of course that's just the mechanical work done at the cranks propelling the bike forward, and not the total energy metabolised, which will be approximately 4-5 times that value depending on a few things, primarily a rider's individual gross mechanical efficiency (GME - the ratio of energy reaching the cranks as a proportion of total energy metabolised). The vast majority of energy we metabolise ends up as waste heat. That's just the warm blooded Mammalian way.

A (good) indicator of energy metabolised
Somewhat serendipitously, since 1 Cal (kcal) ~= 4.2kJ, we can as a reasonable first approximation use the kJ reading from a power meter file (e.g. 700kJ) and make a straight conversion of that number to energy metabolised (e.g. 700 Cal) since the GME and conversion of kJ to Cal (almost) neatly cancel each other out. The real conversion is probably more like in the range of:
1.05 - 1.15 x kJ of mechanical work done  = Calories metabolised.

A measure of fitness
The Average Power a rider can maximally sustain in a well-paced steady state effort such as during a flat time trial or on an indoor trainer is one of the most direct and objective measures of fitness. It is usually expressed in terms of maximal average (mean maximal) power for various durations (e.g. 1-minute, 5-minutes, 1-hour), and in terms of watts per kilogram of body mass (W.kg-1).

It should come as no surprise that we can sustain a higher power output over shorter durations. Over the course of a training block, we seek to raise the power a rider can maximally sustain per kilogram of body mass for durations of relevance to the rider's target events. The higher the mean maximal W.kg-1 number, the faster one can ride and/or the longer a rider can sustain a given pace. Along with a consideration of the specific demands of a rider's events, this is a fundamental principle that should guide a rider's training.

### Normalized Power

So what happens when power output is highly variable, such as typically happens when we ride outdoors over variable terrain, or with a group, in a road, criterium or track race or over a mountain bike course; or perform interval efforts at various power levels with rest periods interspersed?

Racing, group rides, hills all provide for highly variable efforts.

In these common scenarios, Average Power can be a misleading indicator of intensity and understate the level of difficulty of a ride (often substantially so).

That’s because, and to quote Andy Coggan:
1. the physiological responses to rapid changes in exercise intensity are not instantaneous, but follow a predictable time course, and
2. many critical physiological responses (e.g., glycogen utilization, lactate production, stress hormone levels) are curvilinearly, rather than linearly, related to exercise intensity.
This latter point is really important. As power output goes up, the level of strain experienced increases exponentially.

By way of example, let’s say a rider is capable of maximally sustaining 200 watts for about an hour . If we asked them to perform a 20-minute steady paced effort at 200 watts, then assuming they are not unduly fatigued, we should expect the rider could actually complete such an effort, since by definition they are capable of sustaining that power output for longer than 20-minutes. It would be hard, but do-able (indeed, over 20-minutes, a rider could typically maximally sustain ~ 104-109% of their 1-hour power).

But what if we asked the same rider to perform a 20-minute effort with the same average power of 200 watts, except this time the rider is asked to perform 10 x 2-minute interval repeats comprising 300 watts for 1-minute followed by 100 watts for 1-minute?

Those with any experience of this sort of effort will know the rider would be very unlikely to successfully execute the prescribed session, despite the average power being the same. This is because the strain experienced during the 300 watt sections is far greater than the relative increase in power, and is not equally matched by the reduced level of strain experienced when riding the 100 watt '"recovery" sections.

Normalised Power is a clever means by which reported power output is adjusted to take into account the typical and natural variability in power output. To quote Dr Coggan:
“Normalised power provides a better measure of the true physiological demands of a given training session - in essence, it is an estimate of the power that you could have maintained for the same physiological "cost" if your power output had been perfectly constant (e.g., as on a stationary cycle ergometer), rather than variable. Keeping track of normalised power is therefore a more accurate way of quantifying the actual intensity of training sessions, or even races.”
This is one reason why we track Normalised Power, as it represents a more accurate indicator of the level of difficulty and is a helpful guide to changes in fitness over the medium and longer terms when the vast bulk of training data comprises rides of variable effort levels.

Feasible training sessions
Interval training, i.e. the use of periods of higher intensity work coupled with recovery periods, is quite a common feature in many training plans (usually because it can be highly effective in improving fitness). Normalized Power is very helpful in establishing whether a proposed training session is "physiologically feasible".

In the interval example quoted earlier (the 10 x 2-min 300W / 100W intervals), the Normalised Power for such a session would be 234 watts, meaning the equivalent physiological cost of riding at a sustained steady state 234 watts. Typically you would expect a rider with an FTP of 200 watts to be able to maximally sustain ~ 104-109% of their FTP for 20-minutes, or ~ 208-218 watts.

Hence the original prescribed session was unrealistic from the outset. You can use Normalised Power in this manner to guide the level of difficulty of training sessions, so that they are hard enough to provide sufficient stimulus to improve fitness but are not so hard they become impossible to execute. Nifty huh?

The underlying physiological principles and the mathematics of the Normalised Power algorithm are described in more detail in an article by Dr Coggan quoted earlier in this post.

Caveats
There are limitations and caveats to how one uses and interprets Normalized Power, and that's for Part II, so stay tuned....

## Saturday, March 23, 2013

### A time for a bit of sensitivity (analysis)

The Performance Manager Chart is a tool that's been with us for a while, being first released into the wild by Andy Coggan, and the guys from Training Peaks circa 2006. Before then it was tested by a dozen or so lunatics in a power meter users' asylum known as "TSTWKT".

In the years since, it and its off-shoot variants have become a ubiquitous tool for power meter users to inspect the "forest" that represents our overall training loads, as well as giving additional insight into our training patterns and as a indicator of likely form, either prospectively, or as a retrospective analysis tool.

People use the tool as one guide to their overall training progress, to check their actual and planned workload is appropriate for their current training cycle and training objectives. Of course it's only one part of the picture and as always, one must tend to the individual trees, that is, be concerned with the composition of one's training to ensure the specificity principle of training is not lost in the undergrowth.

There's been plenty written about these issues and the use and sometimes misunderstanding of the use of the tool. I'm not going to delve into the whole shebang here, rather just touch upon one small element about the Performance Manager Chart that the more experienced and/or astute user of this tool will understand.

A quick recap:
The basic Performance Manager Chart plots three things - Acute Training Load (ATL), Chronic Training Load (CTL) and Training Stress Balance, where today's TSB = yesterday's (CTL - ATL). It can also show other information if desired, such as daily training stress scores, best power performances and so on.

In layman's terms, ATL is an indicator of how hard you've been training in recent weeks, and CTL is an indicator of how hard you've been training in recent months. ATL and CTL are both exponentially weighted moving averages of the daily Training Stress Scores (TSS), which in turn are calculated from a rider's power meter data and their current threshold power.

Since ATL and CTL are exponentially weighted moving averages, a key input into their calculation is a time constant. The default time constants used for the PMC are 7-days for ATL and 42-days for CTL.

I thought I'd demonstrate with a video animation what happens if you change these defaults settings and comment on whether and/or why you should or would do so. Cue the (94-second long) video:

Occasionally the question asked is - what time constants should I adopt?

The answers usually include the following points:

• Suggest that you create a range of Performance Manager Charts, each with a different combination of time constants, and see which you consider best reflects your actual performances.
• Note that the chart is not particularly sensitive to changes in the CTL time constant, so you may as well leave that at the default 42-day setting.
• The chart is far more sensitive to changes in the ATL time constant, and some have suggested using a longer time constant for older/masters age riders, and a shorter one for younger riders with faster recovery time, although I'm unsure I would necessarily use such as rule of thumb, as there's more to it than just age.
• Even so, changes to the ATL TC (such that one would still consider it an acute indicator) don't radically change the fundamental patterns displayed on the chart, just the absolute values along with a slight time phase shift in the TSB. Keep in mind that it's the patterns that are more insightful than the absolute numbers.
• If you really want to go there, there is software (RaceDay Apollo) and a method described by Dr Phil Skiba to test yourself regularly such that the "ideal" time constants for you can be calculated, although there is likely a sizeable error range in such calculation of ideal time constants and the effort required to do the frequent regular performance testing to narrow that range is likely beyond the training desire of most.
• If you are a multi-sport athlete, then it gets pretty complicated, as the stress scores from different exercise modalities are not linearly additive, nor will they necessarily use the same time constants.

In my opinion, for vast majority of users there really isn't any need to deviate from the default values, as the additional insight to be gained is likely to be fairly limited. That's not to say it doesn't exist but keep in mind that some won't have TSS data for all rides, and/or TSS values that are possibly subject to errors from an incorrect estimation of threshold power (let alone the chosen source of power data).

But by all means this is not meant to dissuade you from playing with the options. Go forth and explore. Or let coach worry about it. We're good at that.

If you want to read more on the Performance Manager, I suggest the following links as starting points:
My Performance Manager Chart by me
Season Review with a Performance Manager Chart by me again
What is the Performance Manager Chart by Hunter Allen
The scientific inspiration for the Performance Manager by Dr Andrew Coggan

## Thursday, March 07, 2013

### More ZO Zen

A follow up to my post the other day about the setting of zero-offset / torque zero on power meters, and how we need to be sceptical about how auto-zero functions operate (if your power meter uses one).

I received feedback to suggest that auto-zero on SRM could not be as bad as I am suggesting, and that it might indeed be more accurate to leave the auto-zero on.

Well it hasn't been my personal experience that the auto-ZO is as reliable as theory might suggest, so I thought I'd do a test when I got the chance. More backyard science.

So today happened to be a lovely day, and I decided to go for a ride. Not a long one mind you because right now I'm about as fit as Harry the Hairy Nosed Wombat, but a long enough ride outdoors under fairly typical riding conditions for me.

For those that know Sydney - the ride went from Annandale, through Stanmore, past Sydney University, Redfern and onto Centennial Park where I did a 20-min "test" and rode back home again. I've done that ride about a bazillion times.

And before leaving I set my Powercontrol VI to use SRM's auto-zero function, and had my phone camera along to take a few snaps, to see what I noticed along the way. Here's the Powercontrol screen showing zero-offset before starting my ride:

The bottom line is the zero-offset value stored by the Powercontrol, and is the value used when calculating (and storing) power values. The middle, larger number, is the "live" zero-offset reading, akin to the offset number shown in the video in my previous post. i.e. if you apply some force to the cranks, that's the number that will fluctuate along with the force being applied. The "Auto" along the top just tells you that the auto-zero function is enabled.

The temperature inside and outside my home was not all that different, and my meter  had about 10-minutes outside before this initial check. This is a different SRM to the one in the video (different bike), although it's the same model of SRM.

So, my starting zero-offset was 409Hz.

After about 15-minutes I'm at some traffic lights near Redfern Oval, so I take the chance to pull over to the side and see what the zero-offset has done.

We can see that at some stage along the way in that initial 15-minutes, auto-zero has reset the zero-offset value to 417Hz, while the actual zero-offset is 407Hz, 2Hz less than when I left home 15-minutes earlier.

Just so that's clear, that incorrect zero-offset value is now being used to calculate all my power numbers. I have no idea when or how many times during that initial 15-minutes of riding the zero-offset was changed, nor what the size of those changes might have been, other than when I stopped to make this check.

OK, so I continue on to the Park and do a 20-minute test effort. Then after that I leave the Park to head back home, stopping on the bikeway alongside Moore Park to do another check. This is what I see:

Auto-zero has set zero-offset to 419Hz, when the actual zero-offset is 409Hz, same as when I left home about 50-minutes earlier.

Continue on home, and this is the final check I made after about 75 minutes of riding:

Auto-zero has set zero-offset to 403Hz, when the actual zero-offset is 407Hz.

So, to summarise in table format:

So, I have four actual zero-offset readings in a 1:15 ride that vary by only 2Hz (and this is pretty typical for my SRMs, i.e. not much drift in zero readings), yet the auto-zero function has reset the zero-offset value with a range spanning 16Hz. And that's just what I know it's done, let alone what I don't know it's done. As Rumsfeld would say, it's a known unknown.

Perhaps now you can see why I don't use auto-zero on my SRM Powercontrol.

The possible impact to my average power on this ride of a 16Hz differential in zero-offset is 3.5% and for my modest 20-minute test effort today, that's 8 watts. I reckon 8 watts is worth knowing about no matter how fast you are. If it were true, that's nearly one second per km in a time trial. But it ain't.

Now I have no idea whether the auto-zero performed better or worse than that on average because we just don't know. We can never know since no power meter keeps a log of zero-offset changes.

As I said in my previous post, such anomalous changes in zero-offset would make some analysis not worth doing (e.g. aero field testing when you are fine tuning equipment and position choices). I don't know about you, but I think a possible 3.5% variance is pretty significant. It's not something you can correct post-hoc either, since there is no record of what and when changes to zero-offset were made (power values are calculated based on the zero-offset value used at the time of recording).

At least with the Powercontrol, you can easily turn off the auto-zero function (just press the "Pro" button on the zero-offset screen), and checking the zero-offset is trivial press of the Mode & Set buttons at same time.

That's far better than having to navigate through various menus to perform one of the most important checks a power meter user needs to make every time they ride, let alone not being able to disable the auto-zero.

Of course YMMV

## Tuesday, March 05, 2013

### Three, Two, One, Zero Offset

Some more backyard science. Well, training room science perhaps.

This one was prompted by occasional power training forum discussions relating to the setting of torque zero on a power meter, and the auto-torque zero feature on some crank based power meters.

It was also prompted by an addition to my training room set up, which now means I am able to view on a computer screen my SRM zero-offset numbers. That's kind of handy as anyone with an SRM Powercontrol knows, the zero-offset screen only stays on long enough to do a check and set the zero-offset, but then reverts back to the main display screen after a short delay, which is fine for its intended purpose. Since I'm doing something unintended, having the zero-offset on permanent display helps.

So for some fun I put my phone in front of the screen to video record my SRM's zero-offset numbers while testing a few things, namely, how the zero-offset numbers vary from unclipped to being clipped into the pedals. How stable was zero-offset when clipped in? When moving a little but still not attempting to put pressure on the pedals? And what happens when I back pedal?

And while this was on an SRM, the issues arising are applicable to all crank based power meters.

This was the result. It's a 3-minute long video.

OK, so my video ed skills ain't quite up to Francis Ford Coppola standards. The white noise you can hear is my fan that I had left running. Summary thoughts are shown in the video.

Just for the record - here are some more thoughts on this subject:

SRM have an auto-torque zero feature on their wireless units. If you are using an SRM Powercontrol, the auto-zero feature can be enabled or disabled. How the auto-zero function operates is not documented in SRM public literature or on their website, so when and how it invokes is a bit of a mystery. as follows:
1. Speed must be > 0.
2. Cadence must be 0 for at least 5s.
3. The Zero offset must not vary by more than +/- 4 Hz.
If all three of these conditions are met, the new zero offset is the average of the values over the 5s.
Thanks to the contributor that updated me on the SRM function from the German manual.

It's my personal experience that it can generate zero-offset values that are way off. I recommend disabling auto-zero and doing manual zero-offset checks (the same as you do with older wired models). Interesting that SRM says it requires speed > 0, as that implies it also requires a speed separate speed sensor for auto-zero to work.

However, if you use a Garmin device as your head unit with your SRM, then you will have no choice as you presently cannot disable the auto-zero function. In my view that's a significant functional flaw that Garmin and SRM should fix. How significant? For example, I would not rely on Garmin captured data from an SRM when performing aerodynamic field tests. Use an SRM Powercontrol.

Quarq does not have an auto-zero function, the user needs to choose to perform a torque zero (which is fine by me, it's far better than having an auto-zero you can't disable and have no control over or knowledge of when it happens). A torque zero can be done manually as normal or by back-pedalling the cranks a sufficient number of rotations (at least four).

Back pedalling to set a torque-zero is convenient for sure but introduces an error similar to that described in the video. The size of that error will vary and depends on how different your individual reading is compared to the fixed back pedal torque value assumed by Quarq. Best to check and set your torque-zero manually, and unclipped from the pedals, and preferably not when coasting either (on many bikes this latter item is no big deal but some have a bit of freehub drag that can apply positive torque to the cranks while coasting).

Power2Max enables you to do a manual torque zero check as normal and it also uses an auto-zero function which you cannot disable (at least not with a Garmin). P2M have publicly stated the auto-zero function will only trigger if the torque readings are stable for a period and presumably the crank is not rotating for a few seconds. The maximum torque variance that would trigger an auto-zero being no more than one "ppm", with "ppm" being the unit the P2M uses for torque measurement (each power meter reports using different units).

Using a filter of stable torque readings makes sense to prevent erroneous torque-zero values but I wonder how often that actually happens given it was not easy for me to keep a stable zero-offset even when on a trainer and able to focus on doing just that.

P2M's reported torque units are about one-quarter to one-fifth as sensitive as those displayed by an SRM. So the torque values a P2M would interpret as being a stable zero point and trigger an auto-zero, would be the equivalent of an SRM zero-offset value being within a range of ~5-10Hz. So while it seems likely that the P2M will trigger an auto zero when coasting with reasonable frequency, the consequence of this level of (in)sensitivity in the torque range used to trigger an auto-zero means it could well introduce a random error of up to +/- 5W in power readings.

Powertaps of course are not subject to the same issue of trying to deal with pedal forces when coasting since they are measuring torque at the freehub, and so an auto-zero can be invoked as the hub will know when it is coasting (and hence no torque is being applied). It's not perfect either, and there are situations when it might be fooled, but in general it works reasonably well.

Note that the Powertap auto-zero feature when using a Powertap Cervo head unit (Little Yellow Computer) will only work if the torque-zero is not too far out of range to begin with (up to 8 Powertap torque units I think but that's from the dark recesses of my memory), so it's important to perform a manual torque-zero check before starting any ride. I don't know if this function is the same when using Garmin head units. Auto-zero on a Powertap can be disabled on both the LYC and Garmin.

Finally, the torque units reported by a Powertap and used to invoke an auto-zero are about one seventh as sensitive as those on an SRM (it depends on the gearing used and the range is typically one-quarter to one-tenth as sensitive as an SRM for an equivalent crank torque), but at least it has the advantage of being isolated from pedal forces by the freehub.

As a general comment, power meter head units really should be recording torque-zero values and keep a log of when and to by how much those torque-zero values have changed. This is important data to enable forensic examination of a power meter's performance and accuracy. At present, the only power meter head unit to record a torque-zero value in its file is an SRM Powercontrol. Even then, it only records the most recently set zero-offset value.

So while we keep seeing a stack of features being implemented in each new generation of head units, the most basic, fundamental and important feature of a power meter, i.e. the quality of the power data, does not always get the attention it deserves.