I was thinking about Wi-Fi and infrared remote controls for TVs and air conditioners today. (Did you know there used to be infrared 802.11 access points too?) Remotes don’t work if you don’t point them at the air conditioner (or TV or whatever), although the infrared light they emit is pretty bright, as bright as a bright flashlight. But it’s modulated at a barely ultrasonic frequency; the first TV remote controls were in fact ultrasonic chimes pinged with a hammer, and bizarrely manufacturers have continued using the same frequencies despite using a completely different medium.
I’m not totally sure, but I’m guessing that these remotes use infrared LEDs with a forward drop of about 1.5 volts, a current of about 40 mA, and an efficiency on the order of 4%. (Typical LEDs have a luminous coefficient around 3%, although modern illumination white LEDs can reach 25%, but I’m assuming that these advances haven’t translated to the infrared LEDs used in remotes.) That would work out to 2.4 mW of light energy, or +3.8 dBm, which is then emitted with an “antenna gain” of about 6 dBi (thus EIRP ≈ 6.2 dBm) spread over a sphere of perhaps two meters radius (50 m² of surface) and detected by a phototransistor of radius perhaps 8 mm (5e-5 m²), so the received signal should be about 60 dB below the EIRP, or about -54 dBm.
Wi-Fi often achieves sustained, consistent data transmission despite signal levels below -70 dBm, sometimes below -100 dBm. GPS receivers detect even weaker signals.
What if you used the spread-spectrum coding-gain approach used by GPS receivers and Wi-Fi receivers, but for local optical communication? You can easily modulate ordinary LEDs at over a megahertz, and now that CRT TVs are dying out, I think there aren’t many sources of megahertz-range optical noise in households. Common illumination and taillight LEDs are typically pulsed, but only on the order of 100 Hz, in order to reduce switching losses; I think their 10,000th harmonic is going to be very weak, precisely because of those switching losses (aside from having a very low duty cycle even if the switching were perfect).
There is still the problem that direct sunlight is 100 kilolux, and even indoor room lighting is 50–500 lux, while the light you’re able to transmit from an infrared LED is also only about 100 lux a few dozen millimeters in front of the LED, maybe 1 lux or less at the sensor. Even if the infrared light coming in an open window is effectively constant over microsecond-scale time periods, it still gives rise to shot noise at your infrared sensor. We usually think of shot noise as going away as signal levels rise, but actually it increases; it’s just that it increases proportional to only the square root of the signal, so the signal increases even more. But in this case the sunlight “signal” carries no information, and an increase in shot noise proportional to its square root could eventually swamp the information-carrying signal.
In the 1000-ms response time of a conventional remote control, you could transmit literally a million bits, giving you the possibility of up to 60 dB of coding gain. (I think? I’m kind of guessing about the limits of coding gain here. I should go back and really understand information theory.)
The key problem with light is that it is easily blocked, and even when multipath transmission can get it around obstacles, it’s heavily attenuated. But, by the same token, there’s no need to share bandwidth or avoid interfering with reserved bandwidth.