Quantum Fourier Transform (QFT)

• Quantum Fourier Transform is a unitary operation that converts state amplitudes to frequency-domain representations, crucial for algorithms such as Shor’s factoring and phase estimation.
• The standard QFT circuit employs Hadamard gates, controlled phase rotations, and final qubit swaps, yielding an O(n²) gate complexity with room for approximation-based optimizations.
• Advanced variants use truncated rotations and recursive decompositions to reduce resource overhead and adapt to hardware constraints, improving both NISQ and fault-tolerant implementations.

The Quantum Fourier Transform (QFT) is the unitary transform on an $N$-dimensional quantum register, usually with $N=2^n$, defined on computational-basis states by

$$F_N\lvert j\rangle=\frac{1}{\sqrt N}\sum_{k=0}^{N-1}e^{2\pi i jk/N}\lvert k\rangle.$$

On a general state $$\lvert\psi\rangle=\sum_{j=0}^{N-1}a_j\lvert j\rangle$$, the output amplitudes are the discrete Fourier transform of the amplitude vector. Because this phase-to-amplitude conversion admits a textbook implementation with polynomial resources in the qubit count, the QFT is a central subroutine in Shor’s factoring algorithm, phase estimation, QFT-based arithmetic, and a range of sensing and signal-processing protocols (Kashani et al., 2022, Ruiz-Perez et al., 2014, Vorobyov et al., 2020).

1. Algebraic definition and structural forms

For $N=2^n$, the QFT is the unitary matrix $F_N$ with entries

$(F_N)_{k,j}=N^{-1/2}e^{2\pi i jk/N},$

so $F_NF_N^\dagger=I_N$ and $F_N^{-1}=F_N^\dagger$. On an arbitrary $n$-qubit state,

$N=2^n$0

This is the exact quantum analogue of the discrete Fourier transform acting on amplitudes rather than on an explicitly stored classical vector (Kashani et al., 2022).

A standard bitwise form identifies an $N=2^n$1-qubit basis state $N=2^n$2 with the integer $N=2^n$3, and rewrites the QFT output as a tensor product of single-qubit phase states, up to qubit reversal: $N=2^n$4 This form underlies both the textbook circuit decomposition and several classical-simulation results for restricted inputs (Lisnichenko et al., 15 Feb 2025, Abbott, 2010).

The full QFT is often separated into a core transform and a final bit-reversal permutation, $N=2^n$5, where $N=2^n$6. That distinction is not merely notational: later entanglement analysis shows that the operator-entanglement profile of the full transform is dominated by the reversal layer, whereas the core circuit $N=2^n$7 has rapidly decaying Schmidt coefficients (Chen et al., 2022).

A matrix-theoretic viewpoint derives the radix-2 QFT by further decomposing the diagonal factors of the FFT matrix factorization into products with Kronecker-product structure. In that formulation, Hadamards and controlled phase rotations arise directly from the decomposition of the discrete Fourier transform matrix, and the circuit-level relation between FFT-style factorization and the QFT becomes explicit (Camps et al., 2020).

2. Textbook circuit and resource profile

The standard $N=2^n$8-qubit circuit applies one Hadamard to each qubit, followed by controlled phase rotations with angles that decrease exponentially with qubit separation, and ends with a qubit-order reversal. With qubits indexed from the most significant bit, the decomposition is: for each $N=2^n$9, apply $$F_N\lvert j\rangle=\frac{1}{\sqrt N}\sum_{k=0}^{N-1}e^{2\pi i jk/N}\lvert k\rangle.$$0 on qubit $$F_N\lvert j\rangle=\frac{1}{\sqrt N}\sum_{k=0}^{N-1}e^{2\pi i jk/N}\lvert k\rangle.$$1; then for $$F_N\lvert j\rangle=\frac{1}{\sqrt N}\sum_{k=0}^{N-1}e^{2\pi i jk/N}\lvert k\rangle.$$2, apply a controlled $$F_N\lvert j\rangle=\frac{1}{\sqrt N}\sum_{k=0}^{N-1}e^{2\pi i jk/N}\lvert k\rangle.$$3 between control $$F_N\lvert j\rangle=\frac{1}{\sqrt N}\sum_{k=0}^{N-1}e^{2\pi i jk/N}\lvert k\rangle.$$4 and target $$F_N\lvert j\rangle=\frac{1}{\sqrt N}\sum_{k=0}^{N-1}e^{2\pi i jk/N}\lvert k\rangle.$$5, where

$$F_N\lvert j\rangle=\frac{1}{\sqrt N}\sum_{k=0}^{N-1}e^{2\pi i jk/N}\lvert k\rangle.$$6

after all such gates, swap qubit $$F_N\lvert j\rangle=\frac{1}{\sqrt N}\sum_{k=0}^{N-1}e^{2\pi i jk/N}\lvert k\rangle.$$7 with qubit $$F_N\lvert j\rangle=\frac{1}{\sqrt N}\sum_{k=0}^{N-1}e^{2\pi i jk/N}\lvert k\rangle.$$8 for $$F_N\lvert j\rangle=\frac{1}{\sqrt N}\sum_{k=0}^{N-1}e^{2\pi i jk/N}\lvert k\rangle.$$9 (Kashani et al., 2022).

The standard gate counts are exact and compact: $$\lvert\psi\rangle=\sum_{j=0}^{N-1}a_j\lvert j\rangle$$0 Hadamards, $$\lvert\psi\rangle=\sum_{j=0}^{N-1}a_j\lvert j\rangle$$1 controlled-phase gates, and $$\lvert\psi\rangle=\sum_{j=0}^{N-1}a_j\lvert j\rangle$$2 SWAPs. In the usual gate-count model this gives $$\lvert\psi\rangle=\sum_{j=0}^{N-1}a_j\lvert j\rangle$$3 two-qubit gates; several sources also describe reduced depth under more aggressive scheduling or approximation, but the quadratic two-qubit count is the canonical reference point (Kashani et al., 2022, Lisnichenko et al., 15 Feb 2025).

The final swap network implements bit reversal rather than Fourier phases. For applications in which the output is immediately measured or passed to logic that tolerates reversed endianness, the swap layer may be omitted and the result interpreted in reversed order. This convention is used explicitly in multidimensional QFT constructions, where each coordinate block can be measured in bit-reversed order instead of physically swapped (Pfeffer, 2023).

On sparse hardware, the logical all-to-all pattern of controlled phases is not the full compilation problem. The physical cost is strongly architecture-dependent because adjacency constraints can force large SWAP overheads. That observation motivates much of the later literature on approximate, mapped, or asymptotically faster QFT kernels.

3. Approximate, optimized, and asymptotically faster variants

The most common approximation is angle truncation: omit all controlled rotations $$\lvert\psi\rangle=\sum_{j=0}^{N-1}a_j\lvert j\rangle$$4 below a chosen threshold. In approximate QFT (AQFT) with depth parameter $$\lvert\psi\rangle=\sum_{j=0}^{N-1}a_j\lvert j\rangle$$5, all controlled-$$\lvert\psi\rangle=\sum_{j=0}^{N-1}a_j\lvert j\rangle$$6 with $$\lvert\psi\rangle=\sum_{j=0}^{N-1}a_j\lvert j\rangle$$7 are dropped. The gate count then falls from $$\lvert\psi\rangle=\sum_{j=0}^{N-1}a_j\lvert j\rangle$$8 to $$\lvert\psi\rangle=\sum_{j=0}^{N-1}a_j\lvert j\rangle$$9, and choosing $N=2^n$0 yields the familiar $N=2^n$1 scaling. Error bounds are stated by choosing $N=2^n$2, so that the total truncation error is $N=2^n$3 (Basili et al., 2021, Nam et al., 2018, Park et al., 2022).

Fault-tolerant cost is dominated by non-Clifford synthesis rather than by raw gate count. A measurement-and-feedforward AQFT based on a reusable phase-gradient resource state achieves $N=2^n$4 T gates, improving over earlier $N=2^n$5 synthesize-each-rotation approaches (Nam et al., 2018). A later fully coherent construction reduces the leading non-Clifford cost further: the approximate QFT circuit of Park and Ahn has T-count $N=2^n$6 and T-depth $N=2^n$7, eliminates the Toffoli contribution present in the previous design, and uses parallel placement of quantum adders to reduce T-depth (Park et al., 2022).

More recent work changes the recursion itself. A two-subset recursive formulation gives an exact QFT with $N=2^n$8 gates using $N=2^n$9 ancillas, and an approximate QFT with $F_N$0 gates using $F_N$1 ancillas. The same paper attributes the improvement to replacing the traditional one-qubit-at-a-time recursion with a split into two contiguous blocks plus a cross-correction unitary (Shah, 19 Jan 2025). For arbitrary transform size $F_N$2, not necessarily a power of two, a quantum analogue of Bluestein’s algorithm factors the exact $F_N$3-point QFT into three diagonal quadratic-phase gates and two radix-2 QFT subcircuits of size $F_N$4 with $F_N$5, retaining $F_N$6 gate complexity on $F_N$7 qubits (Nan-Hong et al., 17 Dec 2025).

A distinct constant-factor optimization interleaves controlled-phase gates with on-the-fly SWAPs, eliminating a dedicated terminal bit-reversal pass. That circuit preserves the asymptotic $F_N$8 gate and $F_N$9 depth scalings, requires no extra ancillas, and reports small Qiskit-level timing improvements for 4-qubit QFT and a 15-qubit HHL instance (Romero et al., 11 Jul 2025).

Variant	Asymptotic resources	Main mechanism
Standard exact QFT	$(F_N)_{k,j}=N^{-1/2}e^{2\pi i jk/N},$0 gates	Hadamards, controlled phases, final swaps
AQFT	$(F_N)_{k,j}=N^{-1/2}e^{2\pi i jk/N},$1 gates	Drop rotations $(F_N)_{k,j}=N^{-1/2}e^{2\pi i jk/N},$2 with $(F_N)_{k,j}=N^{-1/2}e^{2\pi i jk/N},$3
Faster exact QFT	$(F_N)_{k,j}=N^{-1/2}e^{2\pi i jk/N},$4 gates, $(F_N)_{k,j}=N^{-1/2}e^{2\pi i jk/N},$5 ancillas	Two-subset recursion
Faster approximate QFT	$(F_N)_{k,j}=N^{-1/2}e^{2\pi i jk/N},$6 gates, $(F_N)_{k,j}=N^{-1/2}e^{2\pi i jk/N},$7 ancillas	Bucketed recursion and truncated cross-corrections
Arbitrary-size exact QFT	$(F_N)_{k,j}=N^{-1/2}e^{2\pi i jk/N},$8 gates, $(F_N)_{k,j}=N^{-1/2}e^{2\pi i jk/N},$9 qubits	Quantum Bluestein factorization

4. Arithmetic and transform-domain computation

The QFT supplies a phase-encoded representation in which addition and related arithmetic reduce to controlled phase accumulation. In Draper’s QFT adder, one applies QFT to an accumulator register, performs controlled $F_NF_N^\dagger=I_N$0 rotations conditioned on the bits of the addend, and then applies inverse QFT. The standard modulo-$F_NF_N^\dagger=I_N$1 adder has total gate count $F_NF_N^\dagger=I_N$2, depth $F_NF_N^\dagger=I_N$3, and requires no extra ancillas beyond the two $F_NF_N^\dagger=I_N$4-qubit registers. Variants extend the same pattern to non-modular addition, signed addition and subtraction, multi-input addition with gate count $F_NF_N^\dagger=I_N$5, and multiplication with $F_NF_N^\dagger=I_N$6 two-qubit rotations plus two QFTs (Ruiz-Perez et al., 2014).

The same transform-domain mechanism yields weighted arithmetic. Ruiz-Perez and Garcia-Escartin define a generalized controlled phase shift

$F_NF_N^\dagger=I_N$7

and use it to compute a weighted average $F_NF_N^\dagger=I_N$8 modulo $F_NF_N^\dagger=I_N$9. If the average is not integer, the result can be encoded in fixed-point with $F_N^{-1}=F_N^\dagger$0 qubits, and the phase approximation error is bounded by $F_N^{-1}=F_N^\dagger$1. The same paper identifies weighted sums and inner products as relevant to quantum machine learning and QSVM kernel constructions (Ruiz-Perez et al., 2014).

The arithmetic advantages are offset by hardware sensitivity. QFT-based circuits rely on small-angle rotations $F_N^{-1}=F_N^\dagger$2, which are identified as the most error-prone components on near-term devices, whereas ripple-carry adders use only CNOT and Toffoli gates. This trade-off is explicit in comparative assessments of QFT arithmetic: fewer ancillas and compact qubit layouts come at the price of high-fidelity arbitrary-phase control (Ruiz-Perez et al., 2014).

Noise studies on IBM-style superconducting architectures make the approximation trade-off quantitative. For QFT-based addition with 8-qubit targets, $F_N^{-1}=F_N^\dagger$3 inputs—both addends in single basis states—show nearly $F_N^{-1}=F_N^\dagger$4 success for all $F_N^{-1}=F_N^\dagger$5 even up to $F_N^{-1}=F_N^\dagger$6, whereas $F_N^{-1}=F_N^\dagger$7 and $F_N^{-1}=F_N^\dagger$8 superposed inputs are much more sensitive to two-qubit error rates. For QFT-based multiplication with 4-qubit multiplicands and an 8-qubit product register, even $F_N^{-1}=F_N^\dagger$9 multiplication is significantly noise-sensitive at $n$0, and shallow truncation can nearly double success probability relative to full QFT in high-noise regimes. Across these benchmarks, the optimal AQFT depth depends on both machine noise and the number of superposed states; $n$1 is a recurrent rule of thumb rather than a universal optimum (Basili et al., 2021).

5. Accuracy, entanglement, and classical simulability

A recent accuracy analysis isolates three distinct degradation mechanisms in practical QFT use: discretization or aliasing inherited from classical sampling theory, limited resolution in eigenvalue or phase estimation, and finite quantum resources such as qubit count, gate errors, and decoherence. In that framework, phase-estimation precision scales as $n$2, and two eigenphases $n$3 can be distinguished only if $n$4. The same work states a minimal-amplitude condition for observability above aliasing noise,

$n$5

and uses small-scale simulations to illustrate sharp peaks for integer-aligned tones, spectral leakage for non-integer frequencies, and wrap-around aliasing beyond the representable phase interval (Lisnichenko et al., 15 Feb 2025).

The QFT is not uniformly hard to simulate classically. Abbott showed that on computational-basis inputs, and more generally on a restricted family of product states satisfying explicit separability conditions, the QFT can be de-quantized to a deterministic classical $n$6-time procedure that reproduces the tensor-product output factors. The same analysis identifies separability preservation as the decisive condition: once the QFT output becomes entangled, no analogous $n$7 classical simulation is known (Abbott, 2010).

Later work refines this point by separating the core QFT from bit reversal. The full bit-reversed transform has maximal operator entanglement, but that is attributed entirely to the reversal permutation; the core circuit $n$8 has Schmidt coefficients that decay exponentially quickly and can generate only a constant amount of entanglement independent of system size. In matrix-product-operator form, a constant bond dimension—reported as approximately $n$9 for double-precision accuracy—suffices to represent the QFT core, so applying it to a matrix-product state of low bond dimension takes time linear in the number of qubits. Test calculations on structured data vectors of length $N=2^n$00 to $N=2^n$01 produced speedups of one to three orders of magnitude over FFT, whereas random data saturated the bond dimension and removed the advantage (Chen et al., 2022).

These results do not collapse the role of the QFT in exponential-speedup algorithms. The de-quantization and tensor-network results are restricted by input structure: the hard regime is precisely the regime in which the QFT is embedded inside a larger computation with large superposition, entanglement, or both.

6. Compilation on constrained hardware and alternative physical realizations

On contemporary hardware, the abstract QFT circuit must be adapted to connectivity constraints. A domain-specific mapping framework for Google Sycamore, IBM heavy-hex, and surface-code lattice-surgery topologies exploits two properties: relaxed ordering of commuting CPHASE gates, and recursive unit decomposition of the QFT into sub-kernels. The resulting compiler is reported as the first to guarantee linear-depth QFT circuits on those three architectures, and compared with state-of-the-art approaches it can save up to $N=2^n$02 in SWAP gates and $N=2^n$03 in depth (Jin et al., 2024).

A separate line of work addresses arbitrary connected qubit-connectivity graphs. There the compilation problem is reduced to a $N=2^n$04-covering-path optimization and then to a subset-DP over graph vertices. The resulting generic method minimizes CNOT count for QFT circuits on arbitrary connected graphs and gives similar results to existing optimized circuits for the “sun” and “two joint suns” architectures, while producing a circuit with slightly more CNOT gates for linear-nearest-neighbor architectures (Khadiev et al., 10 Oct 2025).

Not all QFT implementations are gate-by-gate qubit circuits. A resonator-based proposal uses two microwave resonators coupled to an ensemble of transmon-like qubits. An unknown multi-qubit state is mapped into the Fock basis of one resonator, the QFT is realized by free evolution under a cross-Kerr interaction, and the result is localized in the second resonator by projective measurement with success probability $N=2^n$05. The protocol reports qualitative feasibility for a $N=2^n$06-dimensional QFT in current superconducting circuits, state-transfer fidelity above $N=2^n$07, and QFT fidelity above $N=2^n$08 when photon loss and Kerr-switch-off errors are controlled (Chen et al., 2019).

These hardware studies collectively shift the cost model away from the idealized $N=2^n$09 textbook circuit. On NISQ and FT platforms alike, routing, scheduling, native two-qubit interactions, and bit-order conventions materially affect the practical form of the QFT kernel.

7. Signal processing, sensing, and scientific data analysis

Because QFT output probabilities encode discrete Fourier components, it can be used directly for spectral analysis once classical data are amplitude-encoded. A note-detection algorithm built around this idea takes a WAV file, truncates or down-samples it to $N=2^n$10 real samples, normalizes the sample vector, initializes the amplitudes with Qiskit’s initialize, applies the $N=2^n$11-qubit QFT, and reads out the dominant indices $N=2^n$12, which map to physical frequencies by $N=2^n$13. The implementation was validated symbolically with SymPy and executed on the Aer simulator and on ibmq_belem; for a proof-of-concept $N=2^n$14 run with $N=2^n$15 shots on an $N=2^n$16 tone, the correct index was recovered with probability above $N=2^n$17, and after frame-rate adjustment the reported frequency was approximately $N=2^n$18, close to $N=2^n$19 (Kashani et al., 2022).

The same tensor-factor structure extends to multidimensional transforms. For an array with $N=2^n$20 elements equally separated along $N=2^n$21 dimensions, the multidimensional QFT factorizes into $N=2^n$22 commuting blocks acting on disjoint qubit subsets, giving complexity $N=2^n$23. A 6-qubit 2D-QFT implemented on ibm_lagos for an $N=2^n$24 array and measured over $N=2^n$25 shots produced four dominant peaks at $N=2^n$26, in agreement with the classical 2D-FFT of the test function (Pfeffer, 2023).

In quantum sensing, the QFT has been demonstrated on a hybrid register consisting of an NV-center electron spin and three nuclear spins. There the QFT runs on the nuclear spins and decodes phases accumulated by the sensor. The reported precision scales approximately as $N=2^n$27, close to Heisenberg scaling and above the standard quantum limit; for eight levels, the paper compares $N=2^n$28 with the SQL estimate $N=2^n$29. The same protocol demultiplexes overlapping NMR signals of two target nuclear spins by mapping their phases to distinct nuclear-register outputs (Vorobyov et al., 2020).

A more data-intensive application replaces FFT subroutines inside the healpy CMB-analysis workflow with QFT. In Aer-based experiments over HEALPix resolutions $N=2^n$30, total QFT runtime including amplitude encoding exceeded FFT runtime at all tested resolutions, but the pure QFT runtime excluding encoding eventually crossed below the FFT curve around $N=2^n$31. The study identifies amplitude encoding as the dominant bottleneck and concludes that classical-to-quantum loading overhead currently obscures the asymptotic advantage of the QFT in CMB map-making (Farsian et al., 20 May 2025).

Across these applications, the operational pattern is consistent: the QFT itself is often not the limiting primitive. State preparation, connectivity, small-angle control fidelity, and measurement statistics typically dominate the end-to-end resource balance, while the transform remains the phase-processing core around which the broader algorithm is organized.