DrMAL: This game is easily my favorite in the match, the one I feel is a perennial classic. Firstly, because of the opening. 10.Na3 is very useful, it leads to black either losing a pawn or having pieces tied up to defend. White's pawn sac starting with 18.dxc6 keeps advantage in tempo while gaining mobility.
When analyzing with aid of a computer, I am amazed how long it takes engines to foresee 23.Nxh6+ as a strong move as late as on move 22 for black. 23.Nf6+ is also a strong move it is initially preferred by Rybka 4.1 for a few minutes on move 23:
[+1.04] d=18 23.Nxh6+ gxh6 24.Qxh6 f6 25.e5 Bxg2 26.exf6 Rxd6 27.Rxd6 (0:05:33) 34349kN
[+0.92] d=18 23.Nf6+ gxf6 24.Qxh6 Nc5 25.Qe3 e5 26.Qxc5 Qxc5 27.Rxc5 Rfe8 28.Rcc1 Re6 29.Nf5 Rxd1+ 30.Rxd1 Bb5 31.Rd2 Rb6 32.Bf1 Bxf1 33.Kxf1 Na2 34.Kg2 Nc1 (0:02:14) 14566kN
Under deep analysis, Rybka 4.1 chooses 23.e5 as even better, but this assumes black will respond to 23.Nxh6+ counterintuitively by not immediately taking back:
[+1.56] d=23 23.e5 Bxg2 24.Kxg2 Nd5 25.Nf6+ Nxf6 26.exf6 Nb4 27.fxg7 Kxg7 28.Rc4 Qa8+ 29.f3 Nd5 30.Rh4 Kf6 31.Rxh6+ Ke7 32.Nf5+ Ke8 33.Qc2 Rd7 (4:04:00) 1675849kN
[+1.43] d=23 23.Nxh6+ Kh7 24.Rc4 gxh6 25.e5 f5 26.Bxc6 Nxc6 27.Rxc6 Nb4 28.Rc4 Qg7 29.Rc5 Nd5 30.Re1 Ra8 31.Rb5 Ra7 32.Rxa5 Rxa5 33.Qxa5 Rb8 34.Qxa4 Rxb2 35.Qd4 Rb4 36.Qd3 Qg4 (1:44:22) 743931kN
Knowing this, I would still play 23.Nxh6+ because 24...gxh6 is so natural it is not likely to be postponed.
This game is also my favorite because, unlike game 1, after the sacrifice white is significantly better no matter how black responds. In game 1 white's sacrifice could have left black ahead. Black's best response was not very difficult and that game is of "blunder" type.
Moreover, the repercussions of 23.Nxh6+ are surprisingly difficult to foresee and counter. It takes Rybka 4.1 over an hour on a quad core computer at move 22 for black to evaluate 23.Nxh6+ as a strong move. In fact, Rybka chooses 22...Rad8 as best early on, then a very close second for many minutes. In light of this, I consider 22...Rad8 as an invisible inaccuracy. No doubt black well considered 23.Nxh6+ when playing 20...h6, hardly a mistake at that point.
For example, suppose black tries to counter with 22...Nc5 to later protect bishop on c6. This looked like a good prophylactic move to me. It also looks safe to Rybka for awhile, but the evaluation score climbs very slowly in white's favor at greater depth:
[+0.94] d=24 23.Rc4 Ncd3 24.Qc3 Rab8 25.Bf1 Rfd8 26.Rxd3 Nxd3 27.Rxc6 Rb3 28.Qd2 a3 29.bxa3 Qd4 30.Ne3 Ne5 31.Qxd4 Nf3+ 32.Kg2 Nxd4 33.Ra6 Rxa3 34.Nec4 Ra1 35.Ra7 f6 36.e5 Nc6 37.Ra6 Nb4 38.Rxa5 (6:11:04) 2569801kN
After 23.Nxh6 the natural move 23...gxh6 is a mistake as Rybka quickly points out:
[+1.06] d=12 23...Kh7 24.Ng4 f6 25.Ne3 Rd7 26.Qe2 Nc7 27.e5 Bxg2 28.Kxg2 fxe5 29.Nec4 Ncd5 30.Ne4 Kg8 31.Qg4 Rf5 (0:00:20) 2313kN
[+1.98] d=12 23...gxh6 24.Qxh6 f6 25.e5 Qg7 26.Qxg7+ Kxg7 27.Bxc6 Nxc6 28.Rxc6 Nb4 29.Rc5 Nd5 30.exf6+ Kxf6 31.Ne4+ Kg6 32.Rxa5 Rb8 33.Rxa4 Rxb2 34.h4 Rc2 35.Rd2 Rxd2 (0:00:26) 2958kN
After 24.Qxh6 black plays 24...f6 the best move, but white will be up at least two pawns with a winning position. After 25.e5 instead of playing 25...Qg7 that ultimately holds black's loss to two pawns, black tries to maintain material with 25...Bxg2 a blunder that loses sooner.