Concepts → Variance and kill probability

Why 2d6+5 beats 3d4+4 against an 11 HP target — even with nearly identical means

The single counterintuitive thing every D&D player and computer-game min-maxer eventually trips over: when your damage roll's mean is below the threshold, higher variance helps you, not hurts you.

The setup

Two damage rolls. Same exact world: an enemy at 11 HP, one swing left. Pick one:

  • 2d6+5 — range 7–17, mean 12, variance ≈ 5.83
  • 3d4+4 — range 7–16, mean 11.5, variance ≈ 3.75

The means differ by half a point. The minimums are identical. The first roll caps slightly higher. Most players would describe these as basically interchangeable.

They're not.

Drag the HP and watch the winner flip. Below the mean, the higher-variance 2d6+5 wins; well above the mean, the lower-variance 3d4+4 takes over (until both rolls run out of range).

At 11 HP, 2d6+5 wins: 72.22% vs 3d4+4's 68.75% (Δ = 3.47%).

2d6+5
min 7 max 17 mean 12.00 12 ±2.42
P(finish at 11 HP) = 72.22% (13/18)
  • 7 2.78%
  • 8 5.56%
  • 9 8.33%
  • 10 11.11%
  • 11 13.89%
  • 12 16.67%
  • 13 13.89%
  • 14 11.11%
  • 15 8.33%
  • 16 5.56%
  • 17 2.78%
3d4+4
min 7 max 16 mean 11.50 23/2 ±1.94
P(finish at 11 HP) = 68.75% (11/16)
  • 7 1.56%
  • 8 4.69%
  • 9 9.38%
  • 10 15.63%
  • 11 18.75%
  • 12 18.75%
  • 13 15.63%
  • 14 9.38%
  • 15 4.69%
  • 16 1.56%
  • 17 0.00%

The number that matters

At 11 HP, the relevant statistic isn't the mean — it's the probability that the roll lands at or above 11. Worked out from the exact rationals:

  • 2d6+5 finishes the target in 13/18 ≈ 72.22% of swings. There are 36 equally-likely outcomes for 2d6; 26 of them add to 6 or more (the threshold for clearing 11 with a +5 modifier).
  • 3d4+4 finishes the target in 11/16 = 68.75% of swings. There are 64 equally-likely outcomes for 3d4; 44 of them add to 7 or more.

That's a 3.5-percentage-point gap, and it goes the opposite direction from what the means suggest. The lower-mean roll wins.

Why

Both means hover just above the threshold (12 and 11.5 vs HP 11). When the mean is close to the threshold, the kill-probability question reduces to: given that the typical roll is around the threshold, what's the probability of landing on the high side?

For two distributions with similar means, the one with greater variance spreads more probability mass both below and above the mean — but only the "above" side counts toward a kill. Spreading 50% of the mass higher is a strict win when 50% above is what you need.

The intuition flips the moment your mean sits comfortably above HP. If you're swinging for 25 average against an 11 HP target, you're going to one-shot it 99% of the time regardless of variance. In that regime you'd actually prefer the lower-variance roll, because there's a remote chance the high-variance one rolls catastrophically low.

Stated as a rule: kill probability is non-decreasing in variance when mean ≤ threshold; non-increasing when mean ≥ threshold. The crossover sits at the mean itself.

Where this matters in practice

  • Tabletop D&D — the default audience and the cleanest example, but the math is rule-system-agnostic.
  • Baldur's Gate 3 weapon picks at low levels: when you're chip-damaging an enemy whose HP sits just above your average per-swing damage, the higher-variance weapon outperforms. Reverses past mid-game once your modifiers push the mean comfortably over the enemy HP.
  • Diablo 4 / Path of Exile 2 "should I crit-fish?" questions: high-crit-chance builds are higher-variance per swing, which helps when each individual swing isn't already overkill, and hurts when it is.
  • Reliable vs nuke build debates more generally: multi-hit cleave (low variance, similar mean) loses to single-big- swing (high variance, similar mean) in the breakpoint-bound regime where exactly one swing has to threshold the kill.

Try it yourself

Or change the HP in any of those URLs (or the input on the page) — drop it to 8 and the math swaps. Both rolls finish the target almost always, and now the lower-variance 3d4+4 has the slight edge (one-shot reliability beats high-roll potential).

Adjacent reading: When crit chance beats base damage — the same lesson dialled into a single weapon's crit profile. Reliable vs nuke builds — the same lesson with two distributions of the same mean and opposite shapes.

Engine: Rust core with exact rational arithmetic, compiled to a Phoenix NIF for the server-rendered pages above and to a Rust library for native iOS / Android. Same outputs everywhere — guaranteed by the corpus the build gates against.