Superconducting Quantum Computer Architecture

Last week, the transmon showed how choosing \(E_J/E_C\) balances charge sensitivity against anharmonicity. But reduced parameter sensitivity does not mean that every error transition disappears. This week asks a deeper question: can we directly design the quantum states, coupling operators, and Hamiltonian terms so that the environment has difficulty causing relaxation? And if the environment cannot readily drive a transition, will that transition also become harder for us to control?

Driving question for this week: By combining a large inductance with a Josephson junction, fluxonium creates a spectrum unlike the transmon's. Which transitions are protected, and which control or readout operations become more difficult? How do we avoid mistaking a single coherence record for an overall architectural advantage?

View the overall progression of the ten T2 questions

Do not begin by memorizing a transition-rate formula. Begin with the system, the environment, and their coupling. The most general weak-coupling form is

\[ \hat H_{\rm tot} =\hat H_{\rm S}(\lambda_0)+\hat H_{\rm B}+\hat H_{\rm SB}, \qquad \hat H_{\rm S}(\lambda_0)\ket n=E_n\ket n. \]

\(\hat H_{\rm S}\) determines the energy levels and eigenstates, while \(\hat H_{\rm B}\) describes the environment. What actually determines whether the environment can change the qubit state is \(\hat H_{\rm SB}\). If a circuit parameter experiences a small perturbation \(\lambda(t)=\lambda_0+\delta\lambda(t)\), expand the Hamiltonian directly:

\[ \begin{aligned} \hat H_{\rm S}(\lambda_0+\delta\lambda) &\simeq \hat H_0 +\delta\lambda(t) \left.\frac{\partial\hat H_{\rm S}}{\partial\lambda}\right|_{\lambda_0} +O(\delta\lambda^2)\\ &\equiv \hat H_0+F(t)\hat O. \end{aligned} \]

Thus \(F(t)\) is not merely an abstract “noise”; it is the actual fluctuation, while \(\hat O=\partial_\lambda\hat H_{\rm S}\) is determined by the circuit Hamiltonian. If the environment is also quantized, write \(\hat H_{\rm SB}=\sum_\alpha\hat O_\alpha\otimes\hat B_\alpha\); in a semiclassical noise model, \(\hat B_\alpha\) is replaced by a stochastic function \(F_\alpha(t)\). Both descriptions retain the same system-side coupling operator.

Step 1: Expand the coupling operator in the energy eigenbasis

\[ \hat O=\sum_{m,n}O_{mn}\ket m\!\bra n, \qquad O_{mn}=\langle m|\hat O|n\rangle. \]

This immediately separates the same perturbation into two physical effects:

\[ \hat V(t)=F(t) \left[ \underbrace{\sum_n O_{nn}\ket n\!\bra n}_{\text{diagonal: changes phase and transition frequency}} + \underbrace{\sum_{m\ne n}O_{mn}\ket m\!\bra n}_{\text{off-diagonal: changes population}} \right]. \]

• Diagonal part: Each level receives a different instantaneous energy shift, accumulating a random phase over time and causing dephasing.
• Off-diagonal part: Only if \(O_{fi}\neq0\) can the perturbation connect \(\ket i\) to \(\ket f\) and cause excitation or relaxation.

Step 2: Derive the transition amplitude in the interaction picture

Define the Bohr frequency \(\omega_{mn}=(E_m-E_n)/\hbar\). In the interaction picture of \(\hat H_0\),

\[ \hat V_I(t)=F(t)\sum_{m,n}O_{mn}e^{i\omega_{mn}t}\ket m\!\bra n. \]

The interaction-picture Schrödinger equation and the first-order term of the Dyson expansion are

\[ i\hbar\frac{d}{dt}\ket{\psi_I(t)}=\hat V_I(t)\ket{\psi_I(t)}, \qquad \hat U_I(t,0)\simeq 1-\frac{i}{\hbar}\int_0^t dt'\,\hat V_I(t'). \]

If the system begins in \(\ket i\), projection onto \(\ket f\) gives the first-order \(i\to f\) amplitude:

\[ c_f^{(1)}(t) =-\frac{i}{\hbar}O_{fi} \int_0^t dt'\,F(t')e^{i\omega_{fi}t'}. \]

The transition probability is therefore

\[ P_{i\to f}(t) =\frac{|O_{fi}|^2}{\hbar^2} \left| \int_0^t dt'\,F(t')e^{i\omega_{fi}t'} \right|^2. \]

This line already gives two necessary conditions. First, symmetry must not force \(O_{fi}\) to zero. Second, \(F(t)\) must have a Fourier component near \(|\omega_{fi}|\). For a finite observation time, the resonance width is approximately \(1/t\); strict energy conservation emerges only gradually in the long-time limit.

Step 3: Derive coherent-drive and stochastic-noise transitions from the same formula

For an intentional drive, let \(F(t)=F_d\cos\omega_dt\). When \(\omega_d\simeq|\omega_{fi}|\) and the rotating-wave approximation holds,

\[ \Omega_{fi}=\frac{F_d|O_{fi}|}{\hbar}, \qquad P_{i\to f}(t)= \frac{\Omega_{fi}^2}{\Omega_{fi}^2+\Delta^2} \sin^2\!\left(\frac{t}{2}\sqrt{\Omega_{fi}^2+\Delta^2}\right), \quad \Delta=\omega_d-|\omega_{fi}|. \]

This shows that the strength of a control transition is set by the drive-operator matrix element. For stationary random noise, define the correlation function and two-sided spectral density:

\[ C_F(\tau)=\langle F(\tau)F(0)\rangle, \qquad S_F(\omega)=\int_{-\infty}^{\infty}d\tau\, e^{i\omega\tau}C_F(\tau). \]

Ensemble-averaging the amplitude gives

\[ \begin{aligned} \big\langle P_{i\to f}(t)\big\rangle &=\frac{|O_{fi}|^2}{\hbar^2} \int_0^t dt_1\int_0^t dt_2\, e^{i\omega_{fi}(t_1-t_2)}C_F(t_2-t_1)\\ &\xrightarrow[t\gg\tau_c]{} t\,\frac{|O_{fi}|^2}{\hbar^2}S_F(\omega_{if}), \qquad \omega_{if}=\frac{E_i-E_f}{\hbar}. \end{aligned} \]

The Fermi-golden-rule transition rate is therefore

\[ \boxed{ \Gamma_{i\to f} =\frac{1}{\hbar^2} \left|\langle f|\hat O|i\rangle\right|^2 S_F(\omega_{if})}. \]

This result assumes weak coupling, a stationary bath, \(t\) longer than the bath correlation time, and negligible higher-order back-action. If a text absorbs \(2\pi\), \(\hbar\), or the density of states into the definition of spectral density, the prefactor may instead appear in forms such as \(2\pi/\hbar\). Always compare the definition of \(S_F\) before comparing the visual form of equations.

Step 4: Obtain \(T_1\) from the upward and downward rates

Let \(\omega_{01}=(E_1-E_0)/\hbar>0\). The same noise channel gives

\[ \Gamma_\downarrow =\frac{|O_{01}|^2}{\hbar^2}S_F(+\omega_{01}), \qquad \Gamma_\uparrow =\frac{|O_{01}|^2}{\hbar^2}S_F(-\omega_{01}), \qquad \frac{1}{T_1}=\Gamma_\downarrow+\Gamma_\uparrow. \]

For a thermal bath, detailed balance gives \(S_F(-\omega)=e^{-\hbar\omega/k_BT}S_F(+\omega)\). At low temperature, usually \(\Gamma_\uparrow\ll\Gamma_\downarrow\), but upward transitions and thermal population cannot be neglected for a low-frequency qubit or a high effective environmental temperature. For multiple uncorrelated channels, the total rate is the sum of the channel rates; correlated channels require the cross-spectral density \(S_{\alpha\beta}(\omega)\). These upward and downward rates use the unsymmetrized quantum-noise spectrum. A symmetrized spectrum loses the direction of energy flow and cannot distinguish the two on its own.

Step 5: Return to diagonal coupling and derive dephasing and the sweet spot

The diagonal part does not directly exchange population; it changes the energy difference between the two levels:

\[ \delta\omega_{01}(t) =\frac{F(t)}{\hbar}\bigl(O_{11}-O_{00}\bigr), \qquad \delta\phi(t)=\int_0^t dt'\,\delta\omega_{01}(t'). \]

If \(F(t)=\delta\lambda(t)\) and \(\hat O=\partial_\lambda\hat H_0\), the Hellmann–Feynman theorem directly connects the diagonal matrix element to frequency sensitivity:

\[ O_{nn}=\left\langle n\left|\frac{\partial\hat H_0}{\partial\lambda}\right|n\right\rangle =\frac{\partial E_n}{\partial\lambda}, \qquad \frac{O_{11}-O_{00}}{\hbar} =\frac{\partial\omega_{01}}{\partial\lambda}. \]

Therefore,

\[ \delta\omega_{01}(t) \simeq \frac{\partial\omega_{01}}{\partial\lambda}\delta\lambda(t) +\frac12\frac{\partial^2\omega_{01}}{\partial\lambda^2}\delta\lambda^2(t)+\cdots. \]

A sweet spot \(\partial_\lambda\omega_{01}=0\) eliminates only first-order frequency fluctuations. Second-order sensitivity, other noise parameters, and finite-frequency relaxation remain. Dephasing measured by Ramsey, echo, or dynamical-decoupling sequences is the low-frequency \(S_\lambda(\omega)\) weighted by the corresponding filter function.

Step 6: Map the abstract \(F(t)\hat O\) back onto a superconducting circuit

Consider a Hamiltonian containing charging, Josephson, and inductive energies:

\[ \hat H_0 =4E_C(\hat n-n_g)^2 -E_J\cos\hat\varphi +\frac12E_L(\hat\varphi-\varphi_{\rm ext})^2. \]

A small parameter fluctuation does not require guessing an extra operator; differentiate directly:

\[ \delta\hat H(t) \simeq \delta n_g(t)\underbrace{\big[-8E_C(\hat n-n_g)\big]}_{\hat O_{n_g}} +\delta\varphi_{\rm ext}(t) \underbrace{\big[-E_L(\hat\varphi-\varphi_{\rm ext})\big]}_{\hat O_{\Phi}} +\delta E_J(t)\underbrace{\big[-\cos\hat\varphi\big]}_{\hat O_{E_J}}. \]

If the physical flux \(\Phi_{\rm ext}\) is used, multiply additionally by \(\partial\varphi_{\rm ext}/\partial\Phi_{\rm ext}=2\pi/\Phi_0\). Likewise, a capacitive microwave drive usually couples through \(\hat n\). The diagonal sensitivity and off-diagonal matrix element must therefore be computed separately for different operators using the same set of eigenstates.

Charge-like noise: \(\hat n\) and offset charge

Charge noise can arise from oxide defects, two-level systems (TLSs), or gate-voltage fluctuations and couples to the Cooper-pair number:

\[ \hat V_{\rm charge}(t) =-8E_C\,\delta n_g(t)(\hat n-n_g) \simeq-8E_C\,\delta n_g(t)\hat n, \qquad \delta n_g=\frac{C_g\delta V_g}{2e}. \]

Low-frequency charge noise often has \(1/f\) behavior. After increasing \(E_J/E_C\) in a transmon, what is exponentially suppressed is the charge dispersion—the dependence of energy on offset charge \(n_g\). This is a flattening of frequency sensitivity and must not be equated with exponential suppression of \(\langle0|\hat n|1\rangle\).

Flux-like noise: low-frequency dephasing and finite-frequency transitions must be separated

Flux noise can arise from a surface-spin bath. Its low-frequency component primarily causes dephasing through frequency sensitivity:

\[ \text{dephasing weight} \propto \left(\frac{\partial\omega_{01}}{\partial\Phi_{\rm ext}}\right)^2 S_\Phi(\omega\approx0). \]

Noise in the same flux channel near \(\omega_{01}\) instead causes transitions through the off-diagonal matrix element:

\[ \Gamma_{1\to0}^{(\Phi)} =\frac{1}{\hbar^2} \left| \left\langle0\left| \frac{\partial\hat H_0}{\partial\Phi_{\rm ext}} \right|1\right\rangle \right|^2 S_\Phi(+\omega_{01}). \]

A sweet spot and relaxation protection are therefore not the same statement. The former controls diagonal sensitivity; the latter also requires examining the finite-frequency spectrum and off-diagonal matrix element.

Complete decision chain

Under the weak-coupling, Markovian exponential-decay approximation, the full derivation can be compressed to:

\[ \boxed{ \text{circuit and bias} \Rightarrow \hat H_0(\lambda) \Rightarrow \{E_n,\ket n\} \Rightarrow \hat O_\lambda=\partial_\lambda\hat H_0 \Rightarrow \begin{cases} O_{nn}\ \text{and}\ S_\lambda(0) &\Rightarrow T_\phi,\\ O_{mn}\ \text{and}\ S_\lambda(\omega_{nm}) &\Rightarrow T_1\ \text{and excitation}, \end{cases} \Rightarrow T_2^{-1}=(2T_1)^{-1}+T_\phi^{-1}.} \]

For an external perturbation to cause a particular transition, three conditions must hold simultaneously: the environment must couple through a definite \(\hat O\); \(\langle f|\hat O|i\rangle\neq0\); and \(S_F\) must have spectral weight at the required Bohr frequency. “Allowed” means only that symmetry does not forbid the transition, not that it is necessarily strong. “Forbidden” means exactly zero; “suppressed” means nonzero but small. Coherence comparisons must also go beyond a single \(T_1\): matrix-element suppression and reduced environmental spectral density have different implications for device scaling.

Question 1 identified two multiplicative factors in a transition: the environmental spectrum and the system matrix element. In Question 2, assume that the environment cannot be eliminated entirely and ask instead whether the wavefunctions, symmetry, and logical encoding of \(\ket0,\ket1\) can be designed so that the dominant noise operator barely sees the logical transition while another controllable operator remains effective.

Environment-induced relaxation and intentional control have the same mathematical skeleton:

\[ \hat V_{\rm noise}(t)=\hat O_{\rm noise}F_{\rm noise}(t), \qquad \Gamma_\downarrow= \frac{1}{\hbar^2} \left|\langle0|\hat O_{\rm noise}|1\rangle\right|^2 S_{\rm noise}(+\omega_{01}), \]

\[ \hat V_{\rm drive}(t)=\hat O_{\rm drive}A\cos\omega_dt, \qquad \Omega_R=\frac{A}{\hbar} \left|\langle0|\hat O_{\rm drive}|1\rangle\right|. \]

The difference between these expressions is not that “noise is bad and drive is good,” but whether they couple through the same physical port and operator. If \(\hat O_{\rm noise}=\hat O_{\rm drive}\), reducing the qubit matrix element simultaneously lengthens \(T_1\) and lowers the Rabi rate. Only when the two have different symmetry, spatial profiles, or frequency dependence can coupling be weak to the environment yet strong to control.

Step 1: Project every operator into the logical subspace

Define the logical projector

\[ \hat P_{\rm L}=\ket0\!\bra0+\ket1\!\bra1. \]

Any Hermitian operator can be decomposed in the logical subspace as

\[ \hat P_{\rm L}\hat O\hat P_{\rm L} =\bar O\,\hat I +\frac{O_{11}-O_{00}}{2}\hat\sigma_z +\operatorname{Re}O_{01}\,\hat\sigma_x -\operatorname{Im}O_{01}\,\hat\sigma_y, \qquad \bar O=\frac{O_{00}+O_{11}}{2}. \]

The physical requirements for state protection can therefore be read directly as:

• The environment has difficulty distinguishing the logical states:\(O_{11}^{\rm noise}-O_{00}^{\rm noise}\approx0\), reducing first-order dephasing.
• The environment has difficulty exchanging logical population:\(O_{01}^{\rm noise}\approx0\), reducing relaxation and excitation.
• Intentional control is not the identity:\(O_{01}^{\rm drive}\neq0\), preserving transverse control; when logical \(Z\) control is needed, the drive's diagonal difference must also be tunable.

The ideal case can be summarized as

\[ \boxed{ \hat P_{\rm L}\hat O_{\rm noise}\hat P_{\rm L} \simeq c\hat I, \qquad \left|\langle0|\hat O_{\rm drive}|1\rangle\right| \neq0.} \]

This condition must be applied separately to every dominant noise channel. Holding for charge noise does not imply that it also holds for flux, critical-current, dielectric-loss, or quasiparticle channels.

Step 2: How does exact symmetry produce a selection rule?

Suppose a parity-type unitary symmetry \(\hat{\mathcal P}\) exists and satisfies \(\hat{\mathcal P}^2=\hat I\) and \([\hat H_0,\hat{\mathcal P}]=0\). The energy eigenstates can be chosen simultaneously as symmetry eigenstates:

\[ \hat{\mathcal P}\ket n=p_n\ket n, \qquad p_n=\pm1. \]

If the coupling operator has parity \(\eta_O=\pm1\) under this symmetry,

\[ \hat{\mathcal P}\hat O\hat{\mathcal P}^{\dagger}=\eta_O\hat O, \]

then the matrix element must satisfy

\[ \begin{aligned} O_{mn} &=\langle m|\hat O|n\rangle\\ &=\langle m|\hat{\mathcal P}^{\dagger} (\hat{\mathcal P}\hat O\hat{\mathcal P}^{\dagger}) \hat{\mathcal P}|n\rangle\\ &=p_mp_n\eta_O\,O_{mn}. \end{aligned} \]

Therefore,

\[ \boxed{p_mp_n\eta_O=-1\quad\Longrightarrow\quad O_{mn}=0.} \]

• If \(\ket m,\ket n\) have opposite parity, an even operator has a zero matrix element while an odd operator can connect the states.
• If the two states have the same parity, an odd operator has a zero matrix element while an even operator may connect them.
• “May” means only that symmetry does not force the matrix element to zero; its actual magnitude still depends on the wavefunctions.

This is the first way to separate noise from control: give \(\hat O_{\rm noise}\) a forbidden parity and \(\hat O_{\rm drive}\) an allowed parity. In the symmetric half-flux fluxonium potential, states have definite parity. Note carefully, however, that \(\hat n\) and the phase operator relative to the symmetry center are both odd operators, so the \(0\leftrightarrow1\) transition between states of opposite parity is parity-allowed. If that matrix element is small, the physical origin is wavefunction or tunneling suppression; it must not be mislabeled parity-forbidden.

Step 3: How does a symmetry-breaking perturbation open an exact selection rule?

If the real device contains a small symmetry-breaking term,

\[ \hat H=\hat H_0+\epsilon\hat W, \qquad [\hat W,\hat{\mathcal P}]\neq0, \]

first-order perturbation theory gives

\[ \ket{\widetilde n} \simeq\ket n+\epsilon\sum_{k\neq n} \frac{\langle k|\hat W|n\rangle}{E_n-E_k}\ket k. \]

A \(\langle0|\hat O|1\rangle\) that was exactly zero becomes \(O(\epsilon)\) through opposite-symmetry admixture, and the corresponding transition rate generally begins at \(O(\epsilon^2)\) because it depends on the squared matrix element. Thus, a claim of “symmetry protection” must also specify how flux detuning, junction mismatch, stray capacitance, offset charge, or control-line asymmetry breaks the symmetry; showing only the ideal Hamiltonian is insufficient.

Step 4: Wavefunction-overlap suppression is not the same as an exact zero

In the representation of generalized coordinate \(x\),

\[ O_{01}=\langle0|\hat O|1\rangle =\int dx\,\psi_0^*(x)\,\hat O(x,-i\partial_x)\,\psi_1(x). \]

If logical wavefunctions are localized in different wells or regions, or have nearly disjoint support in a multidimensional configuration space, the integral for certain local or smooth operators can be extremely small. This suppression may scale exponentially with the barrier action, effective mass, or an energy-scale ratio, but it is usually not exactly zero: finite tunneling, disorder, mode mixing, and the derivative structure of the operator can all leave a residual matrix element.

The two statements must therefore be kept separate:selection-rule forbidden A zero obtained from symmetry algebra; overlap suppressed A small value obtained from an explicit eigenfunction integral. The former is governed primarily by symmetry-breaking perturbations; the latter by parameter scaling and fabrication variation.

A multi-degree-of-freedom or encoded-state structure can also make a dominant fluctuation act in common mode, so that \(\hat P_{\rm L}(\sum_j c_j\hat O_j)\hat P_{\rm L}\simeq c\hat I\). This protects only the specified collective channel; local disorder or differential noise that projects onto \(\sigma_z\) or \(\sigma_{x,y}\) still causes decoherence.

Step 5: A sweet spot is diagonal protection, not a transition selection rule

For parameter noise \(\lambda\), the Hellmann–Feynman relation from Question 1 gives

\[ \frac{\partial\omega_{01}}{\partial\lambda} =\frac{1}{\hbar} \left( \left\langle1\left|\frac{\partial\hat H_0}{\partial\lambda}\right|1\right\rangle - \left\langle0\left|\frac{\partial\hat H_0}{\partial\lambda}\right|0\right\rangle \right). \]

A sweet spot \(\partial_\lambda\omega_{01}=0\) means that the two logical states have the same first-order diagonal response to that fluctuation, making them harder for the environment to distinguish through a frequency shift. It does not require \(\langle0|\partial_\lambda\hat H_0|1\rangle=0\) and therefore does not automatically imply a longer \(T_1\). Conversely, a small off-diagonal matrix element does not guarantee that \(O_{11}-O_{00}\) is small.

Three forms of protection must therefore be distinguished:

1. Sweet-spot protection: \(\partial\omega_{01}/\partial\lambda=0\), primarily reducing dephasing caused by low-frequency parameter fluctuations.
2. Selection-rule protection: Exact symmetry and operator parity force a particular matrix element to be exactly zero.
3. Matrix-element suppression: The transition is not forbidden by symmetry but is suppressed by wavefunction localization or small overlap.

Step 6: When the direct transition is too weak, what control paths remain?

Preserving controllability does not require making the protected \(0\leftrightarrow1\) transition strong. Common methods include:

1. Use a different operator or port: For example, dominant loss may couple through a charge-like operator while control uses a flux-like operator, or vice versa, exploiting operator parity.
2. Use an auxiliary level: Make \(0\leftrightarrow a\) and \(1\leftrightarrow a\) allowed transitions, then implement a logical rotation with a two-tone or Raman process.
3. Temporarily break the protection: Remain at the protected point while idle; during a gate, apply a bias or parametric modulation that temporarily mixes the states, then return to the protected Hamiltonian after the pulse.
4. Read out through virtual transitions: Even when \(O_{01}\) is very small, the differing virtual couplings from \(\ket0\) and \(\ket1\) to higher states can still produce dispersive contrast.

For a detuned auxiliary state \(\ket a\), eliminating \(\ket a\) gives an effective Raman coupling of the form

\[ \Omega_{\rm eff}\sim \frac{\Omega_{0a}\Omega_{a1}}{2\Delta_a}, \]

The costs include auxiliary-state population, AC Stark shifts, additional decay paths, and calibration complexity. Temporarily breaking symmetry is not free either: a control pulse may also open a previously forbidden noise channel.

Dispersive readout can likewise use virtual transitions through higher states without requiring a strong direct qubit transition. Under the rotating-wave and dispersive approximations, it can be represented schematically as

\[ \chi_j\simeq \sum_{kj}\frac{|g_{jk}|^2}{\omega_{kj}-\omega_r}, \qquad \chi=\chi_1-\chi_0, \qquad \omega_{ab}=\frac{E_a-E_b}{\hbar}>0\ \text{for }a>b. \]

This route still requires checking the critical photon number, Purcell decay, near-resonant higher transitions, and measurement-induced leakage.

Step 7: A stronger drive cannot conceal every controllability problem

If \(O_{01}^{\rm drive}\) is very small, increasing amplitude \(A\) can indeed restore \(\Omega_R\), but the same pulse also couples to \(\ket{k\ge2}\). Define the closest leakage detuning from the drive:

\[ \Delta_{\rm leakage} =\min_{i\in\{0,1\},\,k\ge2} \left|\omega_{ki}-\omega_d\right|. \]

For an off-resonant leakage channel, the lowest-order estimate is

\[ P_{\rm leakage} \sim \sum_{i\in\{0,1\},\,k\ge2} \left| \frac{A\langle k|\hat O_{\rm drive}|i\rangle} {\hbar(\omega_{ki}-\omega_d)} \right|^2. \]

A useful state structure must therefore do more than make \(O_{01}^{\rm noise}\) small: it must create a sufficiently large hierarchy between the desired drive matrix element and undesired leakage matrix elements. Direct resonant control generally requires \(\Omega_R\ll\Delta_{\rm leakage}\); an actual gate must also account for the pulse spectrum, AC Stark shift, counter-rotating terms, and calibration error.

Complete design-check sequence

1. Specify the logical states \(\ket0,\ket1\) and every dominant \(\hat O_{\rm noise}^{(\alpha)}\).
2. Compute each projected operator: \(O_{11}-O_{00}\) and \(O_{01}\), evaluating dephasing and transition protection separately.
3. State whether a small matrix element comes from exact symmetry, wavefunction overlap, or merely a smaller environmental \(S(\omega)\).
4. Add realistic asymmetry and determine how a forbidden matrix element scales with mismatch or bias error.
5. Choose \(\hat O_{\rm drive}\) and either a direct or auxiliary transition, then calculate the Rabi rate.
6. List every nearby transition and check leakage, readout, reset, and thermalization—not only idle coherence.

Why not simply continue adjusting the transmon's \(E_J/E_C\)?

Fluxonium can be viewed as a rearrangement of charge- and flux-qubit design. It does not merely substitute a different set of numbers into the transmon Hamiltonian; it adds a very large inductance, creating new global confinement, flux dependence, and a multiwell potential landscape. Compared with the transmon, fluxonium commonly has a lower computational frequency, greater anharmonicity, and a longer relaxation time at selected operating points.

The three principal elements of fluxonium

• Josephson junction: Provides the nonlinear cosine potential and \(E_J\).
• Capacitor: Provides the charging energy \(E_C\).
• Superinductor: Provides the small inductive energy \(E_L\) and phase confinement.

A superinductor must combine very large inductance, low loss, and low parasitic capacitance. It may be implemented with a Josephson-junction array, a high-kinetic-inductance material, or another high-impedance structure. These implementations also introduce array modes, process variation, and parasitic coupling.

Step 1: Write the three circuit-element energies using branch flux

Choose the branch flux across the small junction as the generalized coordinate:

\[ \Phi(t)=\int^t V(t')dt', \qquad \varphi=\frac{2\pi\Phi}{\Phi_0}, \qquad \varphi_{\rm ext}=\frac{2\pi\Phi_{\rm ext}}{\Phi_0}. \]

The capacitor, Josephson junction, and linear inductor respectively contribute

\[ T_C=\frac12C_\Sigma\dot\Phi^2, \qquad U_J=-E_J\cos\varphi, \qquad U_L=\frac{(\Phi-\Phi_{\rm ext})^2}{2L}. \]

The location of external flux in this expression depends on the gauge and coordinate choice. The physical content is that the fluxoid constraint makes the small-junction phase, superinductor phase, and external flux mutually dependent.

Step 2: Perform a Legendre transform on the Lagrangian

The Lagrangian is kinetic energy minus potential energy:

\[ \mathcal L =\frac12C_\Sigma\dot\Phi^2 +E_J\cos\!\left(\frac{2\pi\Phi}{\Phi_0}\right) -\frac{(\Phi-\Phi_{\rm ext})^2}{2L}. \]

The canonical charge conjugate to \(\Phi\) is

\[ Q=\frac{\partial\mathcal L}{\partial\dot\Phi} =C_\Sigma\dot\Phi, \qquad n=\frac{Q}{2e}, \qquad [\hat\Phi,\hat Q]=i\hbar \Longleftrightarrow [\hat\varphi,\hat n]=i. \]

Perform the Legendre transform:

\[ \begin{aligned} \hat H &=\hat Q\dot{\hat\Phi}-\mathcal L\\ &=\frac{\hat Q^2}{2C_\Sigma} -E_J\cos\hat\varphi +\frac{(\hat\Phi-\Phi_{\rm ext})^2}{2L}\\ &=4E_C\hat n^2 -E_J\cos\hat\varphi +\frac12E_L(\hat\varphi-\varphi_{\rm ext})^2, \end{aligned} \]

where

\[ E_C=\frac{e^2}{2C_\Sigma}, \qquad E_L=\frac{1}{L}\left(\frac{\Phi_0}{2\pi}\right)^2. \]

The most direct answer to Question 3 is therefore:A superinductor adds a quadratic inductive-energy term \(\tfrac12E_L(\hat\varphi-\varphi_{\rm ext})^2\). More importantly and counterintuitively, larger \(L\) means smaller \(E_L\). A superinductor supplies global confinement with low curvature that extends across many cosine wells.

Step 3: Two common gauge expressions represent the same Hamiltonian

Define the new coordinate as \(\hat\theta=\hat\varphi-\varphi_{\rm ext}\); the Hamiltonian can be rewritten as

\[ \hat H =4E_C\hat n^2 +\frac12E_L\hat\theta^2 -E_J\cos(\hat\theta+\varphi_{\rm ext}). \]

Thus the two forms commonly seen in the literature

\[ -E_J\cos\hat\varphi +\frac12E_L(\hat\varphi-\varphi_{\rm ext})^2 \qquad\text{and}\qquad -E_J\cos(\hat\theta+\varphi_{\rm ext}) +\frac12E_L\hat\theta^2 \]

differ only by a coordinate shift. A derivation or numerical calculation must use one gauge consistently from beginning to end; the external-flux shift must not be added to both the cosine and the parabola.

Step 4: Why does an inductive shunt change the phase coordinate and offset charge?

A transmon-like circuit can be written as

\[ \hat H_{\rm T}=4E_C(\hat n-n_g)^2-E_J\cos\hat\varphi. \]

Without an inductive term, the potential is \(2\pi\)-periodic, phase is a compact coordinate, and \(n_g\) enters the wavefunction boundary condition. Adding the inductor makes \(\varphi\) an extended coordinate on the full real line: \(\varphi\to\varphi+2\pi\) moves to another cosine well and is no longer the same configuration point.

If offset charge is retained temporarily,

\[ \hat H=4E_C(\hat n-n_g)^2+U(\hat\varphi), \qquad \hat U_g=e^{in_g\hat\varphi}, \qquad \hat U_g^\dagger(\hat n-n_g)\hat U_g=\hat n. \]

Because \(\varphi\in\mathbb R\), this unitary transformation does not introduce the twisted boundary condition of a compact coordinate, and static \(n_g\) can be removed from the low-energy Hamiltonian. This is the structural origin of fluxonium being “free of charge offsets.” It does not mean that every capacitive-loss channel or finite-frequency noise coupled through \(\hat n\) disappears.

Step 5: Why can a Josephson-junction array approximate a superinductor?

In practice, \(N\) larger Josephson junctions are often connected in series. Let the phase drop across each array junction be \(\theta_j\), with \(E_{J,a}\gg E_{C,a}\) so that quantum phase slips in each junction are rare. Expanding for small \(\theta_j\):

\[ -E_{J,a}\cos\theta_j \simeq-E_{J,a}+\frac12E_{J,a}\theta_j^2. \]

If the collective phase drop is \(\Theta=\sum_{j=1}^{N}\theta_j\), the lowest-energy distribution is approximately \(\theta_j=\Theta/N\), so

\[ U_{\rm array} \simeq\sum_{j=1}^{N}\frac12E_{J,a}\left(\frac{\Theta}{N}\right)^2 =\frac12\frac{E_{J,a}}{N}\Theta^2 \equiv\frac12E_L\Theta^2. \]

Therefore, in the simplest uniform-array approximation,

\[ E_L\simeq\frac{E_{J,a}}{N}, \qquad L_{\rm array}\simeq N\frac{(\Phi_0/2\pi)^2}{E_{J,a}}. \]

Adding series junctions increases total inductance and lowers effective \(E_L\). But \(N\) cannot be increased arbitrarily: junction capacitance, island-to-ground capacitance, parameter disorder, and array length together create a family of internal plasma or array modes.

Step 6: What happens mathematically to the potential after the quadratic term is added?

The effective potential is

\[ U(\varphi) =-E_J\cos\varphi +\frac12E_L(\varphi-\varphi_{\rm ext})^2. \]

Local minima \(\varphi_m\) satisfy

\[ E_J\sin\varphi_m +E_L(\varphi_m-\varphi_{\rm ext})=0, \]

The curvature near each minimum and the small-oscillation energy scale are approximately

\[ U''(\varphi_m)=E_J\cos\varphi_m+E_L, \qquad \hbar\omega_{p,m}\simeq \sqrt{8E_C\big(E_J\cos\varphi_m+E_L\big)}. \]

When \(E_J/E_L\) is large, the cosine produces many local wells while the slowly varying parabola sets their global energy ordering. Ignoring local zero-point corrections,

\[ U_m\approx-E_J+\frac12E_L(2\pi m-\varphi_{\rm ext})^2. \]

External flux can therefore reorder different fluxoid-like wells. At \(\varphi_{\rm ext}=\pi\), neighboring wells can become a symmetric degenerate pair. \(E_C\) and the barrier determine interwell tunneling—the wavefunction engineering pursued in Question 4.

Step 7: Under what conditions can an array be reduced to a single \(E_L\)?

A real array is not an ideal coil without dynamics. When all array-island fluxes are retained, the Hamiltonian contains a capacitance matrix and multiple Josephson cosines:

\[ \hat H_{\rm array} =\frac12\hat{\mathbf Q}^{\mathsf T}\mathbf C^{-1}\hat{\mathbf Q} -\sum_{j=1}^{N}E_{J,a}^{(j)}\cos\hat\theta_j. \]

Reducing it to one quadratic term requires at least:

1. Each array junction lies in the approximately linear regime \(E_{J,a}/E_{C,a}\gg1\) and has a sufficiently low quantum phase-slip rate.
2. The lowest internal array-mode frequency lies above the relevant qubit, control, and readout frequency range.
3. Hybridization between array modes and the qubit collective coordinate is sufficiently weak, or is explicitly included in a multimode Hamiltonian.
4. Ground capacitance, junction disorder, dielectric loss, and quasiparticle loss do not turn the superinductor into a new dominant noise source.

Once these conditions fail, “adding a superinductor” no longer means adding only \(\tfrac12E_L\varphi^2\). It also adds parasitic modes, avoided crossings, additional decay and leakage paths, and fabrication sensitivity.

Why not simply replace the superinductor with a large resistor?

The goal of superinductance is zero DC resistance, low dissipation, and an impedance above the Cooper-pair resistance quantum \(R_Q=h/(2e)^2\approx6.45\,\mathrm{k}\Omega\) across the operating band, while pushing self-resonance outside that band. A large resistor may also have high impedance, but it introduces real admittance, Johnson–Nyquist noise, and energy dissipation, and cannot provide the same coherent Hamiltonian term.

• \(\hat n\): the Cooper-pair number operator conjugate to phase \(\hat\varphi\).
• \(\hat\varphi\): the phase difference across the small junction.
• \(\varphi_{\rm ext}=2\pi\Phi_{\rm ext}/\Phi_0\): the phase offset caused by external flux.
• \(E_C\): the charging energy, controlling the charge kinetic term.
• \(E_J\): the Josephson energy, controlling the depth of the cosine wells.
• \(E_L\): the inductive energy, set by the superinductor.

Layout-to-Hamiltonian mapping

Chip design does not merely draw a circuit; it directly sculpts the Hamiltonian's parameter space. At the same time, process variation becomes \(E_J,E_C,E_L\) spread, frequency offsets, and yield problems.

Question 3 derived the fluxonium Hamiltonian from the circuit. Question 4 goes beyond asking “what term was added” and tracks how each energy scale changes the eigenvalue problem:

\[ \hat H =4E_C\hat n^2 -E_J\cos\hat\varphi +\frac12E_L(\hat\varphi-\varphi_{\rm ext})^2, \qquad [\hat\varphi,\hat n]=i. \]

Step 1: Write the Hamiltonian as a phase-basis Schrödinger equation

In the phase basis, \(\hat n=-i\partial_\varphi\), so

\[ \left[ -4E_C\frac{\partial^2}{\partial\varphi^2} +U(\varphi) \right]\psi_k(\varphi) =E_k\psi_k(\varphi), \]

where

\[ U(\varphi)=-E_J\cos\varphi +\frac12E_L(\varphi-\varphi_{\rm ext})^2. \]

Comparing with an ordinary one-dimensional Schrödinger equation, the effective mass of the phase particle is

\[ m_\varphi=\frac{\hbar^2}{8E_C}. \]

Thus \(E_C\) is more than a “charging-energy number”: it sets the quantum kinetic energy and effective mass of the phase particle, while \(E_J,E_L,\varphi_{\rm ext}\) jointly determine the potential landscape it sees.

Competition among three energy scales

Fluxonium's physics comes not from one parameter but from competition among \(E_C,E_J,E_L\). Tuning them simultaneously changes wavefunction localization, tunneling, transition frequencies, anharmonicity, operator matrix elements, and the leakage spectrum.

Step 2: First read localization and zero-point fluctuations within one well

Let \(\varphi_m\) denote a local minimum and define

\[ K_m=U''(\varphi_m)=E_J\cos\varphi_m+E_L. \]

A harmonic expansion around \(\varphi_m\) gives

\[ \hbar\omega_{p,m}\simeq\sqrt{8E_CK_m}, \qquad \varphi_{{\rm zpf},m}=\left(\frac{2E_C}{K_m}\right)^{1/4}, \qquad n_{{\rm zpf},m}=\left(\frac{K_m}{32E_C}\right)^{1/4}. \]

These three quantities provide direct parameter intuition:

• Increase \(E_C\): The phase particle becomes lighter, \(\psi(\varphi)\) broadens, interwell tunneling grows, and the charging kinetic term plays a stronger role in the spectrum.
• Increase \(E_J\): The cosine wells deepen, the barriers rise, and the local curvature increases, but tunneling between wells generally decreases.
• Decrease \(E_L\): The global parabola flattens, the energy detuning between neighboring fluxoid-like wells decreases, and more wells enter the low-energy spectrum.
• Change \(\varphi_{\rm ext}\): This does not directly change the cosine barrier; it shifts the parabola and reorders the relative energies of the wells.

In the deep-well regime, the tunnel splitting can be written schematically as \(\Delta\sim\hbar\omega_p e^{-S}\), where the semiclassical action \(S\) increases with barrier height and effective mass. Larger \(E_J/E_C\) therefore generally localizes the states more strongly and reduces \(\Delta\); the exact value still requires solving the full potential.

Step 3: How does external flux select different wells?

When \(E_J/E_L\gg1\), local minima lie near \(\varphi_m\simeq2\pi m\) and their localized-state energies can first be estimated as

\[ E_m^{\rm loc} \approx -E_J +\frac12E_L(2\pi m-\varphi_{\rm ext})^2 +\frac12\hbar\omega_{p,m}. \]

The slope with respect to external flux corresponds to the persistent current:

\[ I_m=-\frac{\partial E_m^{\rm loc}}{\partial\Phi_{\rm ext}} \approx \frac{2\pi E_L}{\Phi_0}(2\pi m-\varphi_{\rm ext}). \]

When the energies of two localized wells approach each other, tunneling turns the crossing into an avoided crossing. External flux controls well detuning; \(E_J/E_C\) primarily controls the tunnel gap; and \(E_L\) controls both the detuning scale and the persistent-current slope.

Half-flux potential, parity, and wavefunctions

At \(\varphi_{\rm ext}=\pi\), let \(x=\varphi-\pi\). The Hamiltonian becomes

\[ \hat H_{1/2} =-4E_C\frac{\partial^2}{\partial x^2} +E_J\cos x+\frac12E_Lx^2, \qquad U(x)=U(-x). \]

The potential has exact reflection symmetry, so eigenstates can have definite parity: the ground state \(\ket0\) is even and the first excited state \(\ket1\) is odd. The bonding/antibonding picture provides intuition, but a rigorous selection rule follows from symmetry and operator parity.

Retaining only the left and right localized states \(\ket L,\ket R\), the low-energy Hamiltonian can be written as

\[ \hat H_{\rm 2w} =-\frac12\left(\Delta\hat\sigma_x+\epsilon\hat\sigma_z\right), \qquad \epsilon\simeq2I_p\,\delta\Phi, \qquad \delta\Phi=\Phi_{\rm ext}-\frac{\Phi_0}{2}. \]

Its qubit splitting and half-flux eigenstates are

\[ \hbar\omega_{01}=\sqrt{\Delta^2+\epsilon^2}, \qquad \ket0\simeq\frac{\ket L+\ket R}{\sqrt2}, \qquad \ket1\simeq\frac{\ket L-\ket R}{\sqrt2}. \]

Away from half flux, \(|\epsilon|\) increases and the eigenstates gradually localize; at half flux they form even and odd superpositions. This two-well model supplies intuition, but higher plasmon states, other wells, and actual matrix elements must be calculated from the complete Hamiltonian.

• Both \(\hat n\) and \(x=\varphi-\pi\) are odd operators and can in principle connect even and odd states.
• If \(\langle0|\hat n|1\rangle\) or \(\langle0|x|1\rangle\) is small, it is not enough to say “parity-forbidden”; the device-specific wavefunction overlap must be checked.
• Even operators such as \(\hat n^2\) and \(x^2\) can be strictly forbidden by a selection rule between opposite-parity states.

Exact parity follows from \(U(x)=U(-x)\) at half flux. Conditions such as \(E_J\gg E_L\) and \(E_J\gg E_C\) are useful in the deep-well/bonding picture but are not required for parity symmetry. Away from half flux, symmetry is broken and first-order flux sensitivity returns, with a magnitude that depends on device parameters.

Step 4: Derive the flux sweet spot from symmetry and the two-well model

The external-flux coupling operator is

\[ \frac{\partial\hat H}{\partial\Phi_{\rm ext}} =-\frac{2\pi E_L}{\Phi_0} (\hat\varphi-\varphi_{\rm ext}). \]

At half flux, \(x=\varphi-\pi\) is an odd operator, and every parity eigenstate satisfies \(\langle n|x|n\rangle=0\). By the Hellmann–Feynman theorem,

\[ \left.\frac{\partial E_n}{\partial\Phi_{\rm ext}}\right|_{\Phi_0/2}=0, \qquad \boxed{ \left.\frac{\partial\omega_{01}}{\partial\Phi_{\rm ext}} \right|_{\Phi_0/2}=0.} \]

The two-well model gives the same conclusion:

\[ \frac{\partial\omega_{01}}{\partial\Phi_{\rm ext}} =\frac{1}{\hbar} \frac{\epsilon}{\sqrt{\Delta^2+\epsilon^2}} \frac{\partial\epsilon}{\partial\Phi_{\rm ext}} \xrightarrow[\epsilon\to0]{}0. \]

At this operating point, first-order flux sensitivity vanishes and dephasing is dominated by higher-order terms. But a sweet spot does not automatically ensure relaxation protection; relaxation still requires the relevant operator matrix element and \(S_F(\omega_{01})\) to be checked.

Moreover, in the two-well approximation,

\[ \left.\frac{\partial^2\omega_{01}}{\partial\Phi_{\rm ext}^2} \right|_{\Phi_0/2} \simeq\frac{(2I_p)^2}{\hbar\Delta}. \]

Making the tunnel gap \(\Delta\) extremely small can reduce \(\omega_{01}\) and some transition matrix elements, but it increases the second-order curvature at the sweet spot and aggravates thermal-initialization and slow-control problems. Protection remains a multiobjective trade-off.

Step 5: Check each noise operator instead of saying vaguely that “fluxonium is protected”

Fluxonium's low-energy states can occupy different phase wells. If two wavefunctions have little overlap, transition matrix elements of some operators also shrink, reducing the probability of \(\ket1\rightarrow\ket0\) through that noise channel. This is wavefunction engineering and must be named separately from an exact parity selection rule.

For Hamiltonian \(\hat H=4E_C\hat n^2+U(\hat\varphi)\), the commutator provides a very useful exact relation:

\[ [\hat H,\hat\varphi]=-8iE_C\hat n, \qquad \left|\langle0|\hat n|1\rangle\right| =\frac{\hbar\omega_{01}}{8E_C} \left|\langle0|\hat\varphi|1\rangle\right|. \]

This shows that, even when the half-flux \(0\leftrightarrow1\) transition has a sizable phase dipole, low \(\omega_{01}\) still suppresses the charge matrix element. Capacitive/dielectric relaxation can therefore be suppressed, whereas relaxation through a flux-like operator does not automatically receive the same suppression; each must still be multiplied by its own \(S_{\rm noise}(\omega_{01})\).

Step 6: Protection must be verified together with control, readout, and leakage

For capacitive and flux-like drives, the direct Rabi rates respectively take the forms

\[ \Omega_R^{(n)}\propto A_n\left|\langle0|\hat n|1\rangle\right|, \qquad \Omega_R^{(\varphi)}\propto A_\varphi\left|\langle0|x|1\rangle\right|. \]

A small charge matrix element slows a direct capacitive gate. Higher transitions in fluxonium, such as \(1\leftrightarrow2\), may have much larger charge matrix elements, so control and readout can use the operator hierarchy, an auxiliary level, or a different coupling port. The cost is more complicated pulse sequences, frequency allocation, and leakage analysis.

At minimum, compute simultaneously

\[ \alpha=\omega_{12}-\omega_{01}, \qquad \Delta_{\rm leakage} =\min_{i\in\{0,1\},\,k\ge2} |\omega_{ki}-\omega_d|, \qquad \Omega_R\ll\Delta_{\rm leakage}. \]

Fluxonium often has greater anharmonicity but also a rich multilevel spectrum. “Large anharmonicity” cannot replace checking every drive matrix element, near resonance, array mode, and multiphoton leakage channel.

A direct cost of a protected spectrum: low frequency and thermal population

Fluxonium's protected regime often comes with a lower computational frequency. For example, with \(f_{01}=500\,\mathrm{MHz}\), \(hf/k_B\approx24\,\mathrm{mK}\): at \(T=50\,\mathrm{mK}\), \(p_1/p_0\approx e^{-24/50}\approx0.62\) and \(p_1\approx0.38\); at \(T=20\,\mathrm{mK}\), \(p_1/p_0\approx0.30\) and \(p_1\approx0.23\).

Another useful scale is \(f_{01}=800\,\mathrm{MHz}\), for which \(hf/k_B\approx38\,\mathrm{mK}\); at \(20\,\mathrm{mK}\), \(p_1\approx0.13\) (only a thermal-order estimate; the actual value also depends on nonequilibrium photons and effective temperature). A low-frequency transition turns active reset, thermalization, and additional readout transitions into design problems.

Complete validation workflow: from parameters to a protection claim

1. Specify \(E_C,E_J,E_L,\Phi_{\rm ext}\), then find \(U(\varphi)\), the local minima, and the barriers.
2. Diagonalize the complete Hamiltonian and plot \(E_k(\Phi)\) and \(\psi_k(\varphi)\); do not rely only on the two-well picture.
3. Calculate \(\partial_\lambda\omega_{01}\) and the second-order curvature, confirming which parameter the sweet spot protects against.
4. For each dominant noise channel, calculate \(\langle0|\hat O_{\rm noise}|1\rangle\) and the corresponding \(S_{\rm noise}(\omega_{01})\).
5. Calculate the drive/readout operator matrix elements, Rabi rate, dispersive path, and \(\Delta_{\rm leakage}\).
6. Compare \(\omega_{01}\) with the refrigerator and effective temperatures, then include thermal population, reset, and bandwidth constraints.

The first four questions used fluxonium to build the complete chain from noise source and operator to Hamiltonian, eigenstates, and matrix elements. Question 5 treats fluxonium as the first complete case and reverses the direction: If the dominant error operator is specified first, should we change parameters, add degrees of freedom, create a new symmetry, or modify the Josephson element itself?

Four indispensable conditions for a protected qubit

These are not four optional advantages; they form a design specification that every protected-qubit proposal must answer in full:

1. Dephasing: Which fluctuating parameter \(\lambda\) is flattened? Only at one sweet spot, or across a finite parameter range?
2. Relaxation: Which \(\hat O_{\rm noise}\) matrix element is suppressed? By exact symmetry or approximate overlap suppression?
3. Control: If noise coupling is suppressed, which different \(\hat O_{\rm drive}\), transition, or auxiliary level is used for control?
4. Leakage: Does protecting the logical states simultaneously create nearby parasitic states, a dense spectrum, or a new leakage path?

Six steps for reverse-engineering a Hamiltonian from an error target

1. Specify the error: First write the dominant fluctuation \(F(t)\), coupling operator \(\hat O_{\rm noise}\), and relevant frequency range. Separate low-frequency dephasing from relaxation near \(\omega_{01}\).
2. Specify the state structure: Decide whether to use a sweet spot, exact symmetry, wavefunction localization, near degeneracy, or disjoint support in multiple degrees of freedom.
3. Synthesize a candidate Hamiltonian: Choose the parameter regime to change, the energy term or degree of freedom to add, the Josephson Fourier component, or the material-dependent element.
4. Validate protection: Solve for the spectrum and eigenstates and calculate \(\partial\omega_{01}/\partial\lambda\) and \(\langle0|\hat O_{\rm noise}|1\rangle\); a potential diagram alone cannot establish protection.
5. Validate operability: Check \(\hat O_{\rm drive}\), the Rabi rate, readout path, auxiliary transition, \(\Delta_{\rm leakage}\), and reset.
6. Map back to hardware: Map every Hamiltonian term onto a capacitor, junction, superinductor, loop, or material, then list newly introduced loss, disorder, parasitic modes, and fabrication sensitivity.

Advanced topic survey: three directions extending Hamiltonian engineering

The teaching objective here is explicitly to create entry points into the terminology and concepts. Students need not complete a full derivation of every architecture this week, but they must see which layer of the Hamiltonian each direction modifies. When they later encounter a new qubit proposal, they should be able to analyze it independently with the same four criteria.

1. \(0-\pi\) qubit: adding degrees of freedom and disconnected state support

Fluxonium can still be viewed primarily as an effective single-degree-of-freedom problem \(H=H(\varphi,n)\). A \(0-\pi\) circuit promotes the problem to a multimode Hamiltonian, for example \(H=H(\phi,\theta)\). The physical circuit uses two Josephson junctions, two large inductors, and two large capacitors. Its goal is not to find another ordinary sweet spot, but to give the logical states nearly disjoint support in generalized-coordinate space while bringing their logical energies close to degeneracy:

• Dephasing: Near degeneracy reduces some parameter sensitivities.
• Relaxation: Separated logical-state support can suppress local-operator matrix elements.
• Control: A drive path crossing the protected subspaces must be designed, often using auxiliary states or symmetry-breaking control.
• Leakage: The multimode spectrum and parasitic modes must be checked together.
• Engineering challenges: Large superinductance, low stray capacitance, component matching, and disorder-induced mode coupling.

Self-study question: Why can adding one degree of freedom enable protection that is difficult to achieve simultaneously with a single phase coordinate?

2. \(\cos(2\varphi)\) qubit: engineering the Josephson potential directly

An ordinary Josephson junction provides \(-E_J\cos\varphi\). More generally, the Fourier components of the current–phase relation can be written as:

\(\cos\varphi\) corresponds to single-Cooper-pair tunneling; an effective \(\cos(2\varphi)\) may correspond to a two-Cooper-pair process, creating a different periodicity and Cooper-pair parity sectors so that symmetry suppresses some charge-induced transitions. This elevates Hamiltonian engineering from “adding \(L\) and \(C\)” to “directly selecting the Fourier content of the Josephson potential.”

• Dephasing: Check the sensitivity of the logical splitting to offset charge, flux, and component asymmetry.
• Relaxation: Use parity sectors to suppress transitions that change the protected parity.
• Control: If a control pulse must temporarily break parity, evaluate the trade-off between protection and gate speed.
• Leakage: The new periodicity rearranges the spectrum; noncomputational states and unwanted Fourier components must be identified.
• Engineering challenges: How can dominant \(E_2\) be realized while suppressing residual \(E_1\cos\varphi\) and fabrication asymmetry?

Self-study question: Must we accept only the \(-E_J\cos\varphi\) naturally supplied by a junction, or can its Fourier components be engineered?

3. Ferro-transmon: modifying the material also modifies the Hamiltonian

The ferromagnetic Josephson-junction direction does not simply add a linear circuit element. It makes the Josephson energy, or more generally the current–phase relation, depend on a magnetic state. A hybrid ferro-transmon can exploit hysteretic behavior so that magnetic pulses provide approximately digital tuning of qubit frequency. But magnetic degrees of freedom simultaneously introduce new noise sources, including magnetic fluctuations and domain dynamics.

• Dephasing: A digital/hysteretic operating point may reduce some continuous-bias sensitivities, but magnetic-state stability must be measured.
• Relaxation: Using a ferroic material does not automatically produce matrix-element protection; \(\hat O_{\rm noise}\) must still be identified.
• Control: The magnetic state provides a new nonvolatile tuning knob.
• Leakage: Check additional material excitations and levels introduced by the modified current–phase relation.
• Engineering challenges: Magnetic fluctuations, hysteresis reproducibility, material interfaces, and process integration.

Core reminder: Hamiltonian engineering is not necessarily protection engineering; a new control knob may simultaneously introduce a new noise operator.

Four criteria × three directions

Fluxonium's own design trade-offs: a protected spectrum is not free

Evaluating fluxonium with the four criteria requires listing the operating frequency, anharmonicity, sweet spot, dominant noise operator, gate primitive, readout contrast, reset, fabrication complexity, coupler compatibility, and frequency allocation. Comparing only the best \(T_1\) ignores the control stack and yield.

In-class synthesis and exit ticket

1. What is the most important equation or model this week, and under which approximations does it apply?
2. Which apparently local parameter actually crosses the device, control, and system layers?
3. If only one new experiment could be performed, which result would most strongly change the current architecture decision?
4. Choose one term from \(0-\pi\), \(\cos2\varphi\), and ferro-transmon. Write down the Hamiltonian term, symmetry, and dominant noise that you need to clarify next.

Cross-layer assessment: Redraw the “abstract model → real hardware” information-flow diagram from the start of the class. The difference between the two diagrams is the evidence of this week's learning. Completion requires a checkable derivation, relation diagram, comparison table, or provisional decision, together with an explanation of its physical assumptions and engineering costs.

Next week: W4 will begin with a driven two-level system and derive the rotating frame and RWA. Students should carry this week's intuition for the qubit Hamiltonian, matrix elements, and leakage gap into W4 to understand why pulse control is constrained by the device spectrum.

Core and Further Reading

• Fluxonium: Single Cooper-Pair Circuit Free of Charge OffsetsV. E. Manucharyan et al., Science 326, 113–116 (2009).The original papers on fluxonium and the superinductor.
• High Fidelity Two-Qubit Gates on Fluxoniums Using a Tunable CouplerI. N. Moskalenko et al., npj Quantum Information 8, 130 (2022).Fluxonium control, reset, coupling, and multiqubit engineering.
• Introduction to Quantum Electromagnetic CircuitsU. Vool and M. H. Devoret, International Journal of Circuit Theory and Applications 45, 897–934 (2017).A core reading on going from a circuit diagram and Lagrangian to a quantum Hamiltonian.
• Engineering High-Coherence Superconducting QubitsI. Siddiqi, Nature Reviews Materials 6, 875–891 (2021).Design trade-offs among qubit architectures, material loss, and coherence.