Assume we have a serial communication system over a wire. Instead of sending pulses for bits, bits can be encoded in phase of a baseband sinusoid, like baseband QPSK(not modulated over a high frequency, each symbol is a sinusoid and phase changes between bits), Why this is not used in practice?
This is more prone to noise and bitrate is higher.
Obviously, we're not talking about "slow" local serial buses (thinking of, say, a 112 kbd UART connection between a microcontroller and a thermal printer, or a 12 Mb/s USB connection), because, well, there's no need to be smarter (read: more expensive) than "hey, high means 1, low means 0, and here's a bit of framing" if even the worst connection suffices for that.
BPSK like that exists in the form NRZ; in that case, the sine wave is at f=0 Hz, but it's just a special case.
The modulation with a non-zero-frequency sine only brings complexity and no real benefits, far as I can tell:
This is more prone to noise and bitrate is higher.
Only as you choose the carrier high enough to allow for orthogonal I and Q, you get twice the channel usage per unit bandwidth, but that's explicitly not what you wanted to do.
The technical reason this is commonly not done is that your receiver needs to be coherent. You can demodulate most baseband buses simply after recovering the symbol clock – no need to also recover the carrier frequency and carrier frequency phase. So, technically, PAM-4 can be easier to implement than QPSK, and gets the same bit/s/Hz. PAM good enough for Fast, Gigabit Ethernet over copper (PAM-5), 2.5, 5 and 10 Gigabit Ethernet over copper wire (PAM-16), and these are already all long-haul links, where the cost-per-lane can be a bit higher than for most serial buses. In PCI Express, gegenerations 1.0 through 5.0 use NRZ, and 6.0 uses PAM-4 – admittedly, with a lot of equalization, but still, this is incoherent reception, allowing a receiver to work relatively quickly.
This is not really a definition or anything, but the moment you have to modulate onto a carrier wave, your device usually stops being called a "serial transceiver" unit or such, and becomes a Modem ("modulator-demodulator", if you wonder where that term comes from). And there's plenty of wired communications over a single wire (so, technically, "serial", but nobody would call them that) that incorporate phase to transport information. The1980 Bell 201A telephone modem at 600 Bd => 1200 b/s uses QPSK, but was to my knowledge the only telephone modem generation that did that – the channel just doesn't fit that very well; you have an SNR that is better than you need for QPSK on a telephone line, if you care enough to equalize, channel code and map with the same effort as needed for high-rate QPSK transmission. Most post-1980 telephone modems go for 16-QAM or higher modulations.
More or less because it doesn't matter. That said, there are examples of non-baseband (DC-balanced or AC coupled) codes, which are used in various applications. Typically these are where galvanic isolation is desirable (Ethernet for example), or where DC isn't transmitted at all by the medium (like magnetic storage media, or radio as such).
I suppose the signals to/from a hard disk head might be a questionable example, as they're not really modulated for sake of the short transmission line between head and head amplifier, heh. But these have used Manchester coding (FM, MFM) for example, with modern examples using more advanced codes. (10BASE-T Ethernet is also Manchester encoded.)
The ratio of bandwidth to center frequency generally depends on the medium. Wire transmission line generally has excellent bandwidth, including down to DC, so it's fine to use broad/base-band signals, like UART, with it. Likewise, AC-coupled signals like Ethernet use quite wide bandwidth (roughly 100kHz to 100MHz), just cutting out the bottom [DC to 100kHz] to allow transformer coupling so you don't have to worry about hazardous voltages on the cable between sites (ground loop, lightning induced surge, etc.).
For lower quality and longer distance lines (particularly those that might be spliced, run through various environments, subject to corrosion, etc., like phone lines), typically a multi-band approach is used. Frequency-division multiplex (FDM) divides the medium's total bandwidth into channels, which may only be usable up to some bitrate each, but by collecting many together into a single link, total bandwidth can be optimized. This includes modern longish-distance high-speed links like DSL and DOCSIS, and older back-haul links like phone trunking lines (where the channels are individual 3kHz phone lines).
For serial/UART traffic, isolation generally isn't required, and at typical low bitrates, cheap optical isolator devices are sufficient (like with MIDI).
Which, optical isolation can be considered an application of modulation itself: the light, though incoherent, is still amplitude-modulated by the signal, and an incoherent detector (photodiode/transistor) suffices to demodulate it.
Bandwidth is one of the limitations of UART style signals, though, and along with typical signaling standards like RS-485, distance is typically limited to a few km. With repeaters or hubs, line loss can be mitigated, but one may well desire a more effective coding scheme over such distances.
