PWM is horrible. Even if well filtered, the signal level is still derived from your (presumably noisy) digital power supply.
A quality DAC will give you true analog voltage, hopefully derived from a stable low-noise reference voltage. There's still a matter of quantization. Real DACs also have non-ideal performance like differential non-linearity (the steps aren't all perfectly the same voltage difference) and integral non-linearity (their voltage response from 0 to max isn't perfectly ideal straight line). But many excellent DACs exist on the market that get these errors very low, well below -90 dB. The very best DACs achieve about -120 dB noise+distortion performance (which is roughly 1 uV error on a 1V scale). Some microcontrollers have pretty decent 12 bit DACs built in, but at best you could expect errors about -60 to -70 dB.
All digital audio synthesis also can suffer from timing jitter. Even if your DAC is perfect, if you output the samples at the wrong times, the result is imperfection added to the final result. If your output is generated by interrupt-driven software, or even worse by simple polling code, interrupt latency can lead to quite a lot of jitter. Generally you need to use a hardware-based approach to control the timing of the sample rate if you want really good quality results. For the highest quality levels, you might also be concerned about the quality of your system clock.
There is a trick call dithering, commonly used with audio DACs, which improves resolution at lower frequencies. As your low frequency waveform slowly transitions from one DAC step to the next, you rapidly output dither patters alternating between those steps. With low-pass filtering, the result is higher resolution than you could have achieved.
DACs are also commonly used with oversampling techniques. If you generate data much, much faster than necessary (and your DAC has the analog bandwidth), then you can greatly lessen the need for analog filtering to remove the ultrasonic content from the rapid steps between DAC voltage levels.
Of course analog filters add noise (thermal noise in resistors & opamps) and distortion (usually due to non-ideal characteristics in their capacitors).
On the synthesis of a high purity sine wave, you might wish to take a look at this code I wrote some time ago:
It synthesizes a sine wave using Taylor series approximation. The phase angle is a 32 bit unsigned integer and the output is 32 bit amplitude, where at least 25 bits match to an ideal sine wave. I'm pretty sure this sine wave is better than can be reproduced by any DAC that exists with today's technology.
You can also find code in that file which generates pretty good sine waves from a modestly sized table of 16 bit data, using linear interpolation between the table entries.