FeynmanDD: Diagram-Centered Computational Framework

• FeynmanDD is a family of diagram-centered computational frameworks that transforms quantum-circuit analysis into a structured, counting-based evaluation of Feynman path integrals.
• It converts circuits into sum-of-powers expressions using multi-terminal binary decision diagrams, enabling efficient amplitude, probability, and equivalence computations via deterministic counting methods.
• Its applications span quantum circuit simulation, many-body physics, and nonlinear spectroscopy, offering a deterministic alternative to stochastic and tensor-network methods.

FeynmanDD denotes a family of diagram-centered computational ideas whose most specific recent meaning is a BDD-based framework for classical quantum-circuit analysis that operates at the level of Feynman path integrals rather than state vectors or unitaries (Wang et al., 10 Sep 2025). In that formulation, a circuit is converted into a sum-of-powers expression, the phase function is represented by a multi-terminal binary decision diagram, and amplitudes, probabilities, and trace-based equivalence criteria are reduced to counting problems. The supplied literature also uses closely related language for deterministic or compiled treatments of Feynman diagrams in many-body physics, scattering amplitudes, and nonlinear spectroscopy, including tensor-train, computational-graph, functional-DAG, and automated diagram-generation frameworks (Ishida et al., 2024).

1. Scope of the term

The literature supplied here suggests that “FeynmanDD” is not yet a single standardized term. It is used most directly as the title of a decision-diagram method for quantum-circuit simulation and equivalence checking, but it also appears as a suggestive label for broader deterministic, compressed, or database-style treatments of Feynman diagrams in other domains.

Usage in the literature	Domain	Core representation
FeynmanDD (Wang et al., 10 Sep 2025)	Quantum-circuit analysis	Standard and multi-terminal decision diagrams applied to path-integral counting
“FeynmanDD” as a deterministic tensor-based blueprint (Ishida et al., 2024)	Multiorbital electron–phonon diagrams	TCI + QTT/TT representations of diagram integrands
Specialized “FeynmanDD” for nonlinear optical response (Rose et al., 2020)	Nonlinear optical spectroscopies	Automated diagram generator for double-sided Feynman diagrams

In the quantum-circuit setting, the central claim is that useful structure lies in the path integral itself. Instead of compressing quantum states or operators, FeynmanDD converts the circuit to a classical function and then uses decision-diagram counting algorithms as the computational engine (Wang et al., 10 Sep 2025). In the many-body and spectroscopy settings, the analogous emphasis is on deterministic evaluation, graph compilation, or automated enumeration of diagrammatic contributions rather than stochastic sampling or manual diagram bookkeeping (Ishida et al., 2024).

2. Path-integral and decision-diagram formulation

For a circuit $U = U_m \cdots U_2 U_1$, the starting point is the Feynman path-integral expression for an amplitude,

$$\langle x|U|0^n\rangle = \sum_{y_1,\dots,y_{m-1}\in\{0,1\}^n} \langle x|U_m|y_{m-1}\rangle \cdots \langle y_2|U_2|y_1\rangle \langle y_1|U_1|0^n\rangle.$$

FeynmanDD re-expresses such amplitudes in a sum-of-powers form

$$\frac{1}{\sqrt{R}} \sum_{y \in \{0,1\}^k} \omega^{f(x,y)},$$

where $\omega$ is a fixed root of unity, $x$ denotes external variables, $y$ internal path variables, and $f$ is a multilinear polynomial modulo $r$ (Wang et al., 10 Sep 2025).

The construction is gate-set dependent. For $\mathcal{Z}=\{H,Z,CZ,CCZ\}$, every Hadamard introduces a new wire variable, whereas diagonal gates reuse variables. The local contributions are

$$\langle y|H|x\rangle = \frac{1}{\sqrt{2}}(-1)^{xy},$$

with $$\langle x|U|0^n\rangle = \sum_{y_1,\dots,y_{m-1}\in\{0,1\}^n} \langle x|U_m|y_{m-1}\rangle \cdots \langle y_2|U_2|y_1\rangle \langle y_1|U_1|0^n\rangle.$$0, $$\langle x|U|0^n\rangle = \sum_{y_1,\dots,y_{m-1}\in\{0,1\}^n} \langle x|U_m|y_{m-1}\rangle \cdots \langle y_2|U_2|y_1\rangle \langle y_1|U_1|0^n\rangle.$$1, and $$\langle x|U|0^n\rangle = \sum_{y_1,\dots,y_{m-1}\in\{0,1\}^n} \langle x|U_m|y_{m-1}\rangle \cdots \langle y_2|U_2|y_1\rangle \langle y_1|U_1|0^n\rangle.$$2 contributing $$\langle x|U|0^n\rangle = \sum_{y_1,\dots,y_{m-1}\in\{0,1\}^n} \langle x|U_m|y_{m-1}\rangle \cdots \langle y_2|U_2|y_1\rangle \langle y_1|U_1|0^n\rangle.$$3, $$\langle x|U|0^n\rangle = \sum_{y_1,\dots,y_{m-1}\in\{0,1\}^n} \langle x|U_m|y_{m-1}\rangle \cdots \langle y_2|U_2|y_1\rangle \langle y_1|U_1|0^n\rangle.$$4, and $$\langle x|U|0^n\rangle = \sum_{y_1,\dots,y_{m-1}\in\{0,1\}^n} \langle x|U_m|y_{m-1}\rangle \cdots \langle y_2|U_2|y_1\rangle \langle y_1|U_1|0^n\rangle.$$5, respectively. For $$\langle x|U|0^n\rangle = \sum_{y_1,\dots,y_{m-1}\in\{0,1\}^n} \langle x|U_m|y_{m-1}\rangle \cdots \langle y_2|U_2|y_1\rangle \langle y_1|U_1|0^n\rangle.$$6, the phase base becomes $$\langle x|U|0^n\rangle = \sum_{y_1,\dots,y_{m-1}\in\{0,1\}^n} \langle x|U_m|y_{m-1}\rangle \cdots \langle y_2|U_2|y_1\rangle \langle y_1|U_1|0^n\rangle.$$7, and for the Google gate set $$\langle x|U|0^n\rangle = \sum_{y_1,\dots,y_{m-1}\in\{0,1\}^n} \langle x|U_m|y_{m-1}\rangle \cdots \langle y_2|U_2|y_1\rangle \langle y_1|U_1|0^n\rangle.$$8 it becomes $$\langle x|U|0^n\rangle = \sum_{y_1,\dots,y_{m-1}\in\{0,1\}^n} \langle x|U_m|y_{m-1}\rangle \cdots \langle y_2|U_2|y_1\rangle \langle y_1|U_1|0^n\rangle.$$9. The paper’s general theorem states that for circuits over $$\frac{1}{\sqrt{R}} \sum_{y \in \{0,1\}^k} \omega^{f(x,y)},$$0, $$\frac{1}{\sqrt{R}} \sum_{y \in \{0,1\}^k} \omega^{f(x,y)},$$1, or $$\frac{1}{\sqrt{R}} \sum_{y \in \{0,1\}^k} \omega^{f(x,y)},$$2, one can efficiently derive an SOP representation in which $$\frac{1}{\sqrt{R}} \sum_{y \in \{0,1\}^k} \omega^{f(x,y)},$$3 has $$\frac{1}{\sqrt{R}} \sum_{y \in \{0,1\}^k} \omega^{f(x,y)},$$4 terms and degree at most $$\frac{1}{\sqrt{R}} \sum_{y \in \{0,1\}^k} \omega^{f(x,y)},$$5 (Wang et al., 10 Sep 2025).

The key reduction is then purely combinatorial. For

$$\frac{1}{\sqrt{R}} \sum_{y \in \{0,1\}^k} \omega^{f(x,y)},$$6

define

$$\frac{1}{\sqrt{R}} \sum_{y \in \{0,1\}^k} \omega^{f(x,y)},$$7

Then

$$\frac{1}{\sqrt{R}} \sum_{y \in \{0,1\}^k} \omega^{f(x,y)},$$8

Thus strong simulation becomes the problem of computing the value distribution of $$\frac{1}{\sqrt{R}} \sum_{y \in \{0,1\}^k} \omega^{f(x,y)},$$9 modulo $\omega$0, and FeynmanDD uses a multi-terminal BDD or MTBDD to represent $\omega$1 and perform that counting exactly (Wang et al., 10 Sep 2025).

3. Graph parameters, linear rank-width, and the treewidth separation

The 2025 analysis of FeynmanDD gives a graph-theoretic characterization of this compression mechanism. For gate sets such as $\omega$2, the SOP polynomial can be written as

$\omega$3

where $\omega$4 is strictly upper triangular over $\omega$5. The associated symmetric matrix

$\omega$6

defines the variable graph $\omega$7 on the SOP variables. Complexity is then controlled not by treewidth but by the cut-rank structure of this graph, quantified by linear rank-width $\omega$8 (Cheng et al., 8 Oct 2025).

For a fixed variable ordering, if $\omega$9 is the cut rank at level $x$0, the number of nodes at that level in the BDD for a $x$1-circuit is at most $x$2. Consequently, under an ordering of width $x$3,

$x$4

Choosing an ordering of width $x$5 yields a FeynmanDD whose size is exponential in $x$6 rather than in treewidth. The same analysis gives $x$7 for the Google gate set $x$8 (Cheng et al., 8 Oct 2025).

This matters because linear rank-width can be substantially smaller than treewidth. The paper explicitly notes that $x$9 is at most larger than the treewidth by a logarithmic factor, but it can also be dramatically smaller; for the complete graph $y$0,

$y$1

That asymmetry underlies the claim that FeynmanDD breaks the treewidth barrier for certain circuit families. The paper constructs explicit IQP-style families for which FeynmanDD runs in polynomial time because the relevant MTBDDs stay polynomial in size, while tensor-network simulation remains exponential because the corresponding treewidth or contraction complexity is $y$2 (Cheng et al., 8 Oct 2025).

The same work also addresses the gate-set restriction. For circuits with arbitrary single-qubit rotations plus $y$3, Solovay–Kitaev compilation to $y$4 increases the linear rank-width of the variable graph by at most $y$5,

$y$6

so the exponential dependence remains on the entangling backbone rather than on the single-qubit approximation overhead (Cheng et al., 8 Oct 2025).

4. Core algorithms and observed performance

Algorithmically, FeynmanDD has three principal tasks: strong simulation, weak simulation, and equivalence checking. Strong simulation computes amplitudes such as $y$7 by substituting external variables into the SOP, building the MTBDD for the resulting polynomial, counting the numbers $y$8, and reconstructing the complex sum. Weak simulation derives probabilities from a doubled polynomial

$y$9

and then uses sequential conditional probabilities to sample output strings. Equivalence checking reduces to evaluating

$f$0

with equivalence up to global phase characterized by

$f$1

for some integer $f$2 (Wang et al., 10 Sep 2025).

To construct the MTBDD efficiently, the method uses binary synthesis rather than monomial-by-monomial accumulation, and it employs variable-ordering heuristics such as qubit order, gate order, and a tensor-order heuristic. It also uses algebraic simplifications that eliminate variables in special parity-constrained situations. Once the MTBDD is built, counting can be performed in time linear in the MTBDD size, up to the bit-length of the exact counts (Wang et al., 10 Sep 2025).

The empirical results reported in the original FeynmanDD paper are strongest for amplitude computation. On Google-style benchmarks, FeynmanDD computes the zero-to-zero amplitude for the circuit cz/5x5_10, with $f$3 qubits and $f$4 gates, in $f$5s and $f$6MB, while DDSIM, SliQSim, and WCFLOBDD time out under the stated $f$7-hour limit. On GHZ and Bernstein–Vazirani families, it scales to $f$8 qubits, with GHZ amplitude computation reported at about $f$9s and sampling at about $r$0s. On the specially designed linear-network circuits, FeynmanDD remains fast while DDSIM and WCFLOBDD time out or use much larger memory (Wang et al., 10 Sep 2025).

The same experiments also delimit the method’s scope. Sampling is more expensive than amplitude computation because the relevant derived SOPs are larger, and on some Google supremacy circuits Quimb’s tensor-network contraction time is reported as faster than FeynmanDD’s runtime, even though FeynmanDD remains faster than TensorCircuit and clearly dominates on the bounded-$r$1 families studied in the follow-up theory paper (Cheng et al., 8 Oct 2025).

5. Relation to broader diagram-centric frameworks

The broader literature supplied here suggests a wider “FeynmanDD-style” agenda: the replacement of brute-force diagram enumeration or stochastic sampling by compact, reusable representations of diagrammatic structure. In condensed-matter many-body theory, one paper explicitly proposes “something one might call ‘FeynmanDD’ (Feynman Diagrams Deterministically / with Deterministic Decompositions)” for weak-coupling self-energy diagrams in a multiorbital electron–phonon model. There the basic ingredients are Tensor Cross Interpolation and Quantics Tensor Train compression, yielding low-rank TT representations of second-order diagram integrands, exponential resolution in imaginary time, and faster-than-power-law convergence of interpolation error with computational effort (Ishida et al., 2024).

A closely related but distinct development treats Feynman diagrams as computational graphs. That framework uses Dyson–Schwinger and parquet equations to organize high-order diagrams into a shared graph of tensor operations, applies Taylor-mode automatic differentiation for renormalization-related derivatives, and compiles the result to CPU or ML backends. In the three-dimensional uniform electron gas, it is reported to reduce the number of primitive operations for sixth-order self-energy by almost three orders of magnitude relative to a direct sum over diagrams (Hou et al., 2024).

For scattering amplitudes, the functional-DAG framework re-expresses recursive amplitude construction as operations on compact abstract data types. Nodes correspond to off-shell wavefunctions or currents indexed by subsets of external legs, and the DAG acts as a compressed representation of the full forest of Feynman diagrams, with shared subcurrents replacing diagram-by-diagram summation (Ohl, 2023). In nonlinear optical spectroscopy, an automated Diagram Generator plays an analogous role for double-sided Liouville-space pathways: it generates all and only the diagrams consistent with a given phase-discrimination condition and set of pulse-overlap intervals, producing $r$2 third-order rephasing diagrams instead of the $r$3 impulsive-limit ones, and $r$4 fifth-order EEI2D diagrams instead of the $r$5 time-ordered ones when all pulses overlap (Rose et al., 2020).

The same diagram-compilation theme appears in automatic perturbative generators and curated databases. A simple Wick-contraction-based algorithm for automatic Feynman-diagram generation starts directly from the perturbative $r$6-matrix formula and generates symmetry factors locally during contraction, with a short C implementation reported to generate $r$7 two-loop diagrams for $r$8 in about $r$9 s on a normal PC (Xiao et al., 2012). Phi4tools, by contrast, packages a high-order data layer: it compiles $\mathcal{Z}=\{H,Z,CZ,CCZ\}$0-, $\mathcal{Z}=\{H,Z,CZ,CCZ\}$1-, and $\mathcal{Z}=\{H,Z,CZ,CCZ\}$2-point $\mathcal{Z}=\{H,Z,CZ,CCZ\}$3 diagrams in fixed dimension, stores their Nickel indices, symmetry factors, integrands, and numerical values up to order eight, and exposes the result as a Mathematica paclet (Sberveglieri et al., 2023). Taken together, these works suggest a broader research program in which “FeynmanDD” refers less to one codebase than to the use of compact algebraic or combinatorial representations for diagrammatic computation.

6. Limitations, misconceptions, and directions

A common misconception is that FeynmanDD is simply another state-vector decision-diagram simulator. The defining distinction of the 2025 method is that it exploits structure in the path-integral formulation and transforms analysis into a counting problem. It does not primarily compress quantum states or operators; it compresses the phase function of the SOP representation (Wang et al., 10 Sep 2025).

Another misconception is that breaking the treewidth barrier means that all hard circuits become easy. The later analysis is explicit that worst-case linear rank-width can still be $\mathcal{Z}=\{H,Z,CZ,CCZ\}$4, so FeynmanDD remains exponential on sufficiently unstructured families. Its advantage is structural rather than universal: it can outperform all tensor-network-based methods on circuit families where $\mathcal{Z}=\{H,Z,CZ,CCZ\}$5 is small and $\mathcal{Z}=\{H,Z,CZ,CCZ\}$6 is large, but it does not subsume tensor-network simulation as a general-purpose method (Cheng et al., 8 Oct 2025).

Current technical limitations are also concrete. In the original formulation, the method directly supports only gate sets with discrete SOP representations such as $\mathcal{Z}=\{H,Z,CZ,CCZ\}$7, $\mathcal{Z}=\{H,Z,CZ,CCZ\}$8, and $\mathcal{Z}=\{H,Z,CZ,CCZ\}$9; arbitrary single-qubit rotations require compilation. Sampling is more expensive than single-amplitude computation because it requires larger derived SOPs involving both $$\langle y|H|x\rangle = \frac{1}{\sqrt{2}}(-1)^{xy},$$0 and $$\langle y|H|x\rangle = \frac{1}{\sqrt{2}}(-1)^{xy},$$1. Performance is sensitive to variable ordering, and optimal ordering is NP-hard. The framework is formulated for pure-state, noiseless circuits, with no native mixed-state or noise model in the supplied account (Wang et al., 10 Sep 2025).

In the broader deterministic-diagram sense, the limitations are different but conceptually related. The tensor-train many-body blueprint is demonstrated only for weak-coupling self-energies up to second order, with $$\langle y|H|x\rangle = \frac{1}{\sqrt{2}}(-1)^{xy},$$2 in the explicit test case, and it leaves higher-order behavior, real-frequency formulations, and rigorous guarantees for the global pivot search open (Ishida et al., 2024). This suggests that the “FeynmanDD” label, across its current uses, marks a methodological direction: replacing direct diagram summation by structured compilation, compression, and exact or deterministic counting, while leaving the choice of representation—MTBDD, TT/QTT, computational graph, functional DAG, or diagram database—domain dependent.