How much uncertainty survives when two quantum light beams mix?

← All digests

Family 273 · Mathematical physics
Importance 5/5 · Provisional editorial rating
Sources consulted 2026-10-25

1. The question

Mix two independent light beams on a beam splitter and keep one output. How uncertain must that output be, given the uncertainties of the inputs? The answer cannot depend only on their average brightness: highly ordered and highly disordered quantum states can carry the same mean number of photons.

A natural comparison uses thermal light. For each input, find a thermal state with the same entropy, then ask whether mixing these thermal references gives the least possible output entropy. This is the entropy photon-number conjecture.

Family 273 reports the inequality for every finite number of modes and finite mean input energy. The two incoming groups must be independent, but the modes inside either group may be arbitrarily entangled. The manuscript then derives an exact constrained output-entropy minimum for identical thermal attenuators and the classical capacity region of a degraded two-receiver pure-loss broadcast channel.

2. Background and essential definitions

Entropy expressed as a photon number

A bosonic mode is a quantum harmonic oscillator with number states $|0\rangle,|1\rangle,\ldots$. The annihilation operator $a$ removes a photon, and $a^\dagger a$ counts photons. An $n$-mode state $\rho$ has finite mean energy in the manuscript’s sense when

$$\operatorname{Tr}\rho\sum_{j=1}^{n}a_j^\dagger a_j<\infty.$$

Its von Neumann entropy is $S(\rho)=-\operatorname{Tr}\rho\log\rho$. All logarithms here are natural, so information is measured in nats. Divide by $\log2$ to convert nats to bits.

A one-mode thermal state with mean photon number $r\geq0$ has geometric number probabilities and entropy

$$g(r)=(r+1)\log(r+1)-r\log r, \qquad g(0)=0.$$

This function increases continuously from zero to infinity, so its inverse $g^{-1}$ exists. For an arbitrary $n$-mode state define its entropy photon number by

$$N_{\mathrm{ent}}(\rho)=g^{-1}(S(\rho)/n).$$

This is the mean photon number per mode of a product of identical thermal states with the same total entropy. It is generally not the actual mean photon number of $\rho$. A pure coherent state, for example, has zero entropy even when it carries nonzero energy.

A beam splitter as a quantum sum

Let $a_j$ and $b_j$ describe corresponding modes at two independent input ports. A passive beam splitter retains modes

$$c_j=\sqrt\eta\,a_j+\sqrt{1-\eta}\,b_j, \qquad 0\leq\eta\leq1.$$

The other output is discarded. “Passive” means the full mixing preserves total photon number; $\eta$ is the power transmissivity. The joint input is $\rho_A\otimes\rho_B$, while $\rho_A$ or $\rho_B$ need not factor across their own modes.

If each input is a product of identical thermal states, with means $r_A,r_B$, the retained output is thermal with mean $\eta r_A+(1-\eta)r_B$. This exact formula motivates a sharp entropy inequality for general states.

3. Development history and previous partial results

The classical entropy power inequality begins with Shannon’s communication theory and was developed through Fisher-information arguments by Stam and Blachman. It bounds the entropy of a sum of independent random vectors. Beam-splitter mixing is a quantum counterpart of addition, but the appropriate thermal entropy function changes the sharp comparison.

Guha, Erkmen and Shapiro formulated the entropy photon-number inequality and explained how it would imply minimum-output-entropy statements relevant to optical communication. The capacity of a single pure-loss bosonic channel had already been determined; noisy and broadcast settings raised sharper constrained entropy questions.

König and Smith established quantum entropy power bounds using Fisher information and heat evolution. De Palma, Mari and Giovannetti obtained the exponential form for every transmissivity,

$$e^{S(\rho_C)/n}\geq \eta e^{S(\rho_A)/n}+(1-\eta)e^{S(\rho_B)/n}.$$

Subsequent work treated more general mixing and gave rigorous finite-energy heat-flow formulations. These were multimode results permitting entanglement within each independent input group.

The desired photon-number inequality is sharper when input entropies differ. Thermal inputs of unequal entropy attain its proposed bound, whereas the exponential inequality does not reproduce their exact output entropy. At equal input entropies, the already known linear entropy bound gives the required photon-number comparison. The full inequality was also known for independent Gaussian inputs, including Gaussian correlations among their modes.

Minimum-output optimization had additional successes. Giovannetti, Holevo and García-Patrón settled broad unconstrained Gaussian optimizer questions. Fixing a positive input entropy imposes a different constraint. De Palma, Trevisan and Giovannetti proved thermal optimality for one-mode phase-covariant Gaussian channels. Multimode progress covered inputs diagonal in a product basis and, in further work of De Palma, entanglement-breaking parameter regimes. Beigi and Rahimi-Keshari provided another one-mode variational route.

The new manuscript claims the two-input comparison without Gaussian, diagonal, product-across-modes or separability restrictions within either port. This is the enlargement that drives its channel consequences.

4. What the new manuscript claims

The two-input inequality

For every finite $n\geq1$, independent finite-energy inputs $\rho_A,\rho_B$, and $0\leq\eta\leq1$, the retained state $\rho_C$ is claimed to satisfy

$$g^{-1}\!\left(\frac{S(\rho_C)}n\right) \geq \eta g^{-1}\!\left(\frac{S(\rho_A)}n\right) +(1-\eta)g^{-1}\!\left(\frac{S(\rho_B)}n\right).$$

Thermal products with a common mean within each port attain equality, including unequal means between the ports. The statement gives these equality examples, not a complete classification of every possible equality case.

The finite-energy assumption ensures the entropies are finite and supports the infinite-dimensional limiting arguments. A finite number of modes does not mean a finite-dimensional Hilbert space: each oscillator can contain arbitrarily many photons.

An exact constrained minimum

A thermal attenuator $\mathcal E_{\eta,N_B}$ mixes one input mode with an independent thermal environment of mean $N_B$. Its $n$-fold tensor power uses identical parameters and a product thermal environment. For finite $s,N_B\geq0$, the manuscript obtains

$$\min_{\substack{\rho\ \mathrm{finite\ energy}\\S(\rho)=ns}} \frac1nS\!\left(\mathcal E_{\eta,N_B}^{\otimes n}(\rho)\right) =g\!\left(\eta g^{-1}(s)+(1-\eta)N_B\right).$$

The input may be entangled across all $n$ modes. A product thermal input of mean $g^{-1}(s)$ attains the minimum. This is attainment, not uniqueness. Vacuum cases and $\eta=0,1$ are included.

Broadcasting to two receivers

One input mode per use is split with independent vacuum auxiliary modes into outputs for receivers $B,C$ and an inaccessible loss output. Assume

$$0\leq\eta_C<\eta_B, \qquad \eta_B+\eta_C\leq1.$$

The weaker receiver’s marginal can be obtained by further attenuating the stronger one’s, which is the meaning of “degraded.” It does not say the physical output systems are independent.

The sender transmits two independent classical messages. Each receiver may collectively measure its own outputs over many uses, with vanishing average decoding error. There is no feedback or receiver cooperation. Codewords have finite energy and may be entangled across uses. The mean photon budget $\bar N$ is averaged over the full codebook and all uses, not imposed separately on every codeword.

The claimed capacity region is the closed convex hull of nonnegative pairs obeying, for some $0\leq\alpha\leq1$,

$$\begin{aligned} R_B&\leq g(\eta_B\alpha\bar N),\\ R_C&\leq g(\eta_C\bar N)-g(\eta_C\alpha\bar N). \end{aligned}$$

Rates are in nats per use. The parameter $\alpha$ divides the coding resource between the two messages; taking the convex hull allows time sharing between choices. If the weaker output is vacuum, its rate is zero and the formula reduces to the known single-receiver expression.

5. Why the advance matters

Importance: 5/5, provisional. The inequality is a central proposed sharp relation between entropy and passive quantum mixing. Its importance includes both the unrestricted multimode input structure and exact communication consequences.

Entanglement across channel uses often prevents a one-use optimization from settling a capacity question. Here the claimed entropy minimum already permits arbitrary finite-block entanglement. For the specified identical thermal attenuators, such correlations cannot lower the constrained output entropy below the thermal benchmark.

The broadcast consequence also gives an operational meaning to the inequality’s sharpness. A bound that misses the thermal output for unequal entropies may leave a gap between achievable rates and a converse. The photon-number form supplies the exact comparison needed to close that gap in the stated degraded pure-loss setting.

The rating concerns mathematical significance if the manuscript’s claims hold; it does not represent independent proof verification.

6. The main idea

Turn a violation into an optimization problem

The manuscript argues by contradiction. Suppose two inputs strictly violate the proposed inequality. Their entropy levels determine thermal reference parameters, which are then held fixed. This converts the violation into a negative value of a functional involving the two input entropies and the output entropy.

Working directly with that functional is delicate. Entropy derivatives can be singular, the Hilbert spaces are infinite dimensional, and creation or annihilation operators are unbounded. The proof therefore adds several regularizations: an auxiliary thermal port, a small thermal replacement at the output, a relative-entropy penalty and a fourth-moment penalty.

These additions serve different purposes. They control energy, make minimizers sufficiently regular, bound output logarithms and justify the traces used later. The argument does not simply assume that entropy can be differentiated along an arbitrary unbounded evolution.

Use curvature adapted to thermal relaxation

At a regularized minimizer, varying one input at a time gives a comparison involving the entropy Hessian, its second-order response. This yields a contraction in a logarithmic-mean metric at those minimizers. The manuscript does not assert that this particular contraction holds universally for every channel and every state.

An independent interpolation theorem combines four comparisons between positive quadratic forms. Its conclusion controls a centered thermal-balance defect: a quantity measuring how creation and annihilation fail to balance as they would in a thermal state, after subtracting the state’s mean.

Pairing the stationarity identities with relaxation generators cancels the centered energy terms. As the fourth-moment penalty is removed, the remaining defects and means vanish. Their number-basis identities force the limiting inputs to be thermal. Entropy continuity then contradicts the strictly negative value that started the construction.

The thermal states are thus reference points and forced limiting optimizers, rather than assumptions about the original inputs. The distinction is what allows non-Gaussian and internally entangled states into the theorem.

Convert entropy into communication bounds

The attenuator minimum follows by choosing the second port to be its thermal environment. For broadcasting, choose a vacuum environment and apply the finite-block bound to conditional states at the stronger receiver. Degradation then bounds the weaker receiver’s conditional entropy. Standard information inequalities supply the converse, following the earlier Guha–Shapiro–Erkmen framework; coherent-state superposition coding supplies achievability.

7. Scope, limitations, and open questions

Independence is required between the two input ports. Arbitrary entanglement within either port is allowed, but correlations across the ports change the problem. The theorem is stated for finite mean energy and finitely many modes; it does not automatically extend to infinite-energy states or infinitely many modes.

The constrained minimum concerns identical passive thermal attenuators. It is not asserted here for mode-dependent transmissivities, amplifiers or additive-noise channels. Thermal attainment does not settle uniqueness of optimizers.

The broadcast region uses independent messages, average energy and error constraints, separate receivers and no feedback. Cooperative decoding, other noise models and different constraints require separate analyses. A full equality classification and extensions beyond the stated finite-energy setting are further questions, rather than consequences assumed in this digest.

8. Sources and further reading

The original entropy photon-number proposal explains the thermal comparison. The earlier quantum entropy power inequality and one-mode constrained optimizer theorem supply key partial results. Guha, Shapiro and Erkmen’s broadcast analysis provides the communication framework. The source below supplies the new claims and precise hypotheses.

Source manuscripts

• M1. OpenAI, The entropy photon-number inequality (September 24, 2026).

Sources consulted on 25 October 2026 at repository revision 1581d4c. Download the manuscript citations.