Phase Gates (S and T)

The $S$/$T$ distinction is not a textbook nicety — it sets the budget of every serious fault-tolerant proposal. In the surface code that Google and IBM are scaling, Clifford gates including $S$ are realized by lattice surgery and Pauli-frame updates that are cheap and essentially error-limited only by the code itself, while every logical $T$ must be grown in a magic-state factory: a large block of the chip dedicated to distilling clean $|T\rangle$ states that are then teleported into the computation. Because those factories dominate both the qubit footprint and the runtime, quantum resource estimates are quoted in $T$-count (or the closely related Toffoli count). The landmark 2021 estimate by Gidney and Ekerå for breaking 2048-bit RSA — roughly 20 million noisy physical qubits running for about 8 hours — is, at bottom, a count of how many magic states Shor's algorithm consumes and how fast the factories can supply them. Compiler research chases the same target from the software side: number-theoretic synthesizers such as gridsynth and the Ross–Selinger algorithm compile arbitrary rotations into near-optimal $H,T$ sequences of length $\approx3\log_2(1/\varepsilon)$, and $T$-count optimizers reshuffle circuits to spend fewer of them. The frontier is making magic cheaper: 2024 "magic-state cultivation" proposals aim to prepare high-fidelity $|T\rangle$ states with far less overhead than traditional distillation, and small logical demonstrations of magic-state injection have begun on superconducting and trapped-ion hardware. Every one of these efforts is, in the end, a war on the $T$-count.

Give a single qubit's relative phase a name: the longitude of its Bloch vector. The two phase gates do exactly one thing to that arrow — they turn its longitude by a fixed amount and leave its latitude untouched. $S$ turns it a quarter of the way around the equator, $90°$; $T$ turns it half of that, $45°$. A state sitting at a pole, $|0\rangle$ or $|1\rangle$, has no longitude to turn and comes back unchanged (up to a global phase no measurement can see); a state on the equator like $|+\rangle$ is swept bodily around. The first figure looks straight down the $z$-axis so the whole motion is a rotation in the plane.

The names are a description of the algebra. $T$ is a smaller turn than $S$, and two $T$s make an $S$; four $T$s make a $Z$, the $180°$ phase-flip; eight $T$s bring the arrow all the way home. So the diagonal phase gates are not a continuum but a discrete clock of $z$-rotations with $T$ as its finest standard tick — the eighth roots of unity laid out on a dial. The second figure draws that clock and colours it by the property that matters most: whether a tick is Clifford or not.

The Clifford/non-Clifford split, which the colours in that clock already anticipate, is the practical heart of the subject. $S$ keeps you inside the stabilizer formalism — a world where every operator is a Pauli and a classical machine can track the full quantum state with a compact tableau. The instant you apply a $T$, conjugating a Pauli stops yielding a Pauli, the state leaves the stabilizer polytope, and you have bought something a classical simulator can no longer follow cheaply. That is the sense in which $T$ is "magic," and it is why $T$, alone among the standard gates, has to be manufactured through distillation. Geometrically the same fact powers universality: a $45°$ $z$-rotation interleaved with the basis-changing Hadamard wanders $SU(2)$ ergodically, so any single-qubit unitary can be approximated arbitrarily well — the content of the Solovay–Kitaev theorem.

Everything starts from two $2\times2$ matrices. Fixing the computational ordering $|0\rangle=(1,0)^{T}$, $|1\rangle=(0,1)^{T}$ turns each gate into four numbers, and because both are diagonal the numbers are trivial to multiply and to power.

where: $i$ — imaginary unit, $i=e^{i\pi/2}$, the relative phase applied by $S$; $e^{i\pi/4}=\tfrac{1+i}{\sqrt2}$ — the relative phase applied by $T$. Read it off directly: both matrices are unitary and diagonal, so they preserve populations and act only on the phase of the $|1\rangle$ amplitude. Applying the matrices to the basis kets shows why "phase gate" is the exact name: the $|0\rangle$ amplitude is never touched, only the $|1\rangle$ amplitude is rotated in the complex plane.

where: $|0\rangle,|1\rangle$ — computational basis states. The reading: because only the relative phase between $|0\rangle$ and $|1\rangle$ changes, the gates are invisible on the poles and maximal on the equator, exactly the picture of the first figure. The nesting relations are the algebra of the clock. Multiplying diagonal matrices adds the phases on the $|1\rangle$ entry, so powering $T$ simply advances one tick at a time.

where: $Z=\mathrm{diag}(1,-1)$ — the Pauli-$Z$ phase-flip; $S^\dagger,T^\dagger$ — the Hermitian conjugates, which are the inverse rotations and equal high powers of the gate itself. The reading: $S$ has order 4 and $T$ order 8, so the inverse of each is just another power — a fact that matters when you assemble phase-kickback circuits and must not confuse $T$ with $T^\dagger$. Both gates are $z$-rotations up to a global phase, and writing them in the symmetric (traceless-generator) form exposes where the misleading "$\pi/8$" name comes from.

where: $R_z(\phi)$ — rotation by angle $\phi$ about the Bloch $z$-axis; $e^{i\pi/4},e^{i\pi/8}$ — the global phases relating $S,T$ to the traceless rotation. The reading: the relative phase $T$ applies is $\pi/4$, but the symmetric generator carries $\pm\pi/8$ on its diagonal — that $\pi/8$, not the physical phase, is the origin of the historical name "$\pi/8$ gate." The Clifford property is the pivot of the whole article, and it is one line: conjugating each Pauli by $S$ returns a Pauli.

where: $X,Y,Z$ — the Pauli operators. The reading: $S$ normalizes the Pauli group — it permutes $\{X,Y,Z\}$ with signs, rotating the $X$–$Y$ plane by $90°$ while fixing the $Z$-axis — so by definition $S$ is a Clifford gate, and any circuit of $H,S,\mathrm{CNOT}$ can be tracked classically. $T$ fails the same test, and the failure is the definition of "non-Clifford." Conjugating $X$ by $T$ rotates it only halfway toward $Y$, landing on an equal mixture that is no Pauli.

where: $\tfrac{1}{\sqrt2}(X+Y)$ — a non-Pauli operator, the $45°$ image of $X$ in the $X$–$Y$ plane. The reading: $T$ does not normalize the Pauli group, so it lies outside the Clifford group; this single fact is what separates classically simulable circuits from universal ones and forces $T$ to be supplied by distillation. Adding $T$ to the Clifford generators makes the gate set universal, and Solovay–Kitaev bounds how expensive that approximation is.

where: $\varepsilon$ — the operator-norm accuracy to which a target single-qubit unitary is approximated; $c$ — the Solovay–Kitaev exponent, provably below 4 and improvable to $c\approx1$ with number-theoretic synthesis. The reading: any unitary is reachable to precision $\varepsilon$ with a number of $H,T$ gates growing only polylogarithmically in $1/\varepsilon$, so universality is not just possible but efficient.

The phase gates are the $\theta=\pi/4$ and $\theta=\pi/2$ members of the diagonal family $\mathrm{diag}(1,e^{i\theta})$; sliding $\theta$ interpolates between familiar landmarks.

1 · The magic state and its measurement statistics (fully worked)

Prepare the $T$-type magic state $|T\rangle=T|+\rangle$ and measure it in the $X$-basis $\{|+\rangle,|-\rangle\}$. Find the state, the outcome probabilities, and its Bloch vector — the numbers a distillation protocol has to improve.

Step 1. Apply $T$ to $|+\rangle=\tfrac{1}{\sqrt2}(|0\rangle+|1\rangle)$: since $T$ fixes $|0\rangle$ and phases $|1\rangle$, the result is $|T\rangle=\tfrac{1}{\sqrt2}(|0\rangle+e^{i\pi/4}|1\rangle)$. This is a state on the equator at longitude $45°$ — exactly the amber point in the first figure.

Step 2. The $X$-basis amplitude is $\langle+|T\rangle=\tfrac12(1+e^{i\pi/4})$, so the probability of outcome $+$ is $|\langle+|T\rangle|^2=\tfrac14|1+e^{i\pi/4}|^2=\tfrac14(2+2\cos\tfrac{\pi}{4})=\tfrac{2+\sqrt2}{4}$. Numerically that is $\mathbf{0.8536}$, and it equals $\cos^2(\pi/8)$; the other outcome carries $\sin^2(\pi/8)=\mathbf{0.1464}$. P(+) = cos²(π/8) ≈ 0.854, P(−) ≈ 0.146.

Step 3. The Bloch vector reads $\langle X\rangle=\cos(\pi/4)=\tfrac{1}{\sqrt2}\approx0.707$, $\langle Y\rangle=\sin(\pi/4)=\tfrac{1}{\sqrt2}\approx0.707$, $\langle Z\rangle=0$: the state points along $(1,1,0)/\sqrt2$. That direction is a midpoint between the $X$ and $Y$ axes — as far as any state can sit from the six vertices of the stabilizer octahedron — which is the geometric meaning of "maximally magic." |T⟩ points to (1,1,0)/√2, the most non-stabilizer equatorial direction.

2 · What one logical T gate actually costs (fully worked)

A fault-tolerant compiler must realize an arbitrary $R_z(\theta)$ rotation to accuracy $\varepsilon=10^{-3}$, and each $T$ it uses must first be distilled from noisy raw magic states of error rate $p=10^{-2}$ using the 15-to-1 protocol. Estimate the $T$-count of the rotation and the distillation rounds needed per $T$.

Step 1. Number-theoretic synthesis (Ross–Selinger / gridsynth) approximates a single $R_z(\theta)$ with a $T$-count of about $3\log_2(1/\varepsilon)$. For $\varepsilon=10^{-3}$ this is $3\log_2(10^{3})=3\times9.97\approx\mathbf{30}$ $T$ gates for one rotation; tightening to $\varepsilon=10^{-10}$ would need $3\log_2(10^{10})\approx100$. ≈ 30 T gates per rotation at ε = 10⁻³.

Step 2. Each of those $T$s is spent as a distilled magic state. One round of the 15-to-1 Bravyi–Kitaev protocol consumes 15 raw states of error $p$ and outputs one of error $p_{\text{out}}\approx35\,p^{3}$. Starting from $p=10^{-2}$: $p_{\text{out}}\approx35\times(10^{-2})^{3}=35\times10^{-6}=\mathbf{3.5\times10^{-5}}$. One round: 10⁻² → 3.5×10⁻⁵.

Step 3. If the algorithm demands logical $T$ error near $10^{-12}$, feed the first output into a second round: $35\times(3.5\times10^{-5})^{3}\approx35\times4.3\times10^{-14}\approx\mathbf{1.5\times10^{-12}}$. Two rounds therefore consume $15\times15=225$ raw states to yield one clean $T$. Two rounds: 225 raw states → one T at ≈ 10⁻¹² error. Multiply by the $\approx30$ $T$s of Step 1 and a single accurate rotation already costs thousands of raw magic states — which is why $S$ (Clifford, free) and $T$ (distilled, dear) sit on opposite sides of every resource budget.

The phase gates are elementary, but the framework that makes $T$ special — the Clifford group, its hierarchy, and the magic states that feed it — was built largely by one person turning quantum error correction into an algebra.

Texts. Nielsen & Chuang, Quantum Computation and Quantum Information, §4.2 (single-qubit gates) and §10.6 (fault tolerance and the $\pi/8$ gate) · Preskill, Lecture Notes on Quantum Computation, Ch. 5–7 · Watrous, The Theory of Quantum Information, Ch. 2 · Wilde, Quantum Information Theory, 2nd ed., Ch. 3 · Kitaev, Shen & Vyalyi, Classical and Quantum Computation (the Solovay–Kitaev theorem) · Gottesman, Stabilizer Codes and Quantum Error Correction (PhD thesis, 1997) · Bravyi & Kitaev, "Universal quantum computation with ideal Clifford gates and noisy ancillas," Phys. Rev. A 71, 022316 (2005). All matrices and identities are verifiable by direct $2\times2$ multiplication: $S=\mathrm{diag}(1,i)$, $T=\mathrm{diag}(1,e^{i\pi/4})$, $T^2=S$, $S^2=Z$, $T^8=I$; the Clifford relation $SXS^\dagger=Y$ and the non-Clifford result $TXT^\dagger=\tfrac{1}{\sqrt2}(X+Y)$. The rotation forms $S=e^{i\pi/4}R_z(\pi/2)$, $T=e^{i\pi/8}R_z(\pi/4)$ follow from $R_z(\phi)=\mathrm{diag}(e^{-i\phi/2},e^{i\phi/2})$. The magic-state statistics $P(+)=\cos^2(\pi/8)=\tfrac{2+\sqrt2}{4}\approx0.854$ are elementary; the synthesis scaling $\approx3\log_2(1/\varepsilon)$ is the Ross–Selinger result and the Solovay–Kitaev bound $O(\log^c(1/\varepsilon))$ with $c<4$. The 15-to-1 distillation output $p_{\text{out}}\approx35p^3$ is the Bravyi–Kitaev protocol; the 2048-bit RSA figures (≈20 million qubits, ≈8 hours) are from Gidney & Ekerå (2021).

Prerequisites. Pauli Gates (X, Y, Z) · Single-Qubit Rotation Gates · Bloch Sphere Representation of Gates · The Quantum Circuit Model. Leads to. Universal Gate Sets · The Controlled-Z Gate · Z-Y Decomposition of Single-Qubit Gates · Circuit Identities and Equivalences.

Ask the tutor: ask me to conjugate any Pauli by $S$ or $T$, walk the order-8 phase clock, or estimate the $T$-count and magic-state cost of a rotation you care about.