---
title: Classical Identification Capacity
url: https://www.emergentmind.com/topics/classical-identification-capacity
type: topic
---

# Classical Identification Capacity

Searching arXiv for recent and foundational papers on identification capacity, including classical channels, deterministic identification, Poisson/Gaussian models, and quantum-channel converse results.
Classical identification capacity is the asymptotic rate at which a channel can support reliable **yes/no tests** of message identity rather than full message recovery. In the identification paradigm, the sender selects a message \(i\), but the receiver’s task is not to decode \(i\) exactly; instead, for each hypothesis \(j\), the receiver wants to decide whether \(j\) was sent. This change in objective produces a different asymptotic theory from ordinary transmission. With randomized encoding, the maximal number of identifiable messages can grow doubly exponentially in blocklength, so the relevant rate is \(\frac{1}{n}\log\log N\). With deterministic encoding, the characteristic growth for continuous-input settings is instead \(2^{R n \log n}\), leading to the rate scale \(\frac{1}{n\log n}\log N\) [2310.16507] [2402.09117].

## 1. Core definitions and asymptotic regimes

For a memoryless channel \(W\), a randomized identification code assigns to each message \(i\) an encoding distribution \(P_i\) on \(n\)-letter inputs and a test region \(E_i\subset \mathcal Y^n\). The two error criteria are the type-I error, requiring \((P_iW^n)(E_i)\ge 1-\lambda_1\), and the type-II error, requiring \((P_iW^n)(E_j)\le \lambda_2\) for all \(j\neq i\). In this regime, one defines capacity through
\[
C_{ID}(W)=\inf_{\lambda_1,\lambda_2>0\atop \lambda_1+\lambda_2<1}\liminf_{n\to\infty}\frac{1}{n}\log\log N_{(n,\lambda_1,\lambda_2)}(W),
\]
or equivalently through the achievability of code sizes \(N=2^{2^{nR}}\) for every \(\lambda\in(0,\tfrac12)\) and all sufficiently large \(n\) [2310.16507] [2606.05032].

Deterministic identification restricts each message to a single codeword rather than an encoding distribution. In that setting, the doubly exponential scale typically disappears. For channels with finite output and arbitrary input alphabet, the relevant asymptotic quantity is
\[
\dot C_{DI}(W):=\inf_{\lambda_1,\lambda_2>0}\liminf_{n\to\infty}\frac{1}{n\log n}\log N_{DI}(n,\lambda_1,\lambda_2),
\]
reflecting the super-exponential but sub-doubly-exponential growth \(2^{\Theta(n\log n)}\) [2402.09117]. For colored Gaussian channels with deterministic encoding under peak power, the corresponding definition is formulated through code sizes
\[
M(n)=2^{(n\log n)R+o(n\log n)}
\]
and the supremum of achievable \(R\) [2604.04674].

A central conceptual distinction follows immediately. In Shannon transmission, reliable communication concerns exact decoding and scales as \(2^{nC}\). In identification, the size of the hypothesis set can be much larger, but the appropriate rate notion changes with the encoding model. This suggests that “more messages” and “larger capacity” are not interchangeable statements.

## 2. Randomized identification and equality with transmission capacity

For discrete memoryless channels, the classical benchmark is that randomized identification capacity equals Shannon capacity. The Poisson-channel analysis makes this analogy explicit: Ahlswede–Dueck and Han–Verdú showed for any DMC that randomized ID capacity equals Shannon capacity, and the discrete-time Poisson channel (DTPC) obeys the same principle despite its continuous input alphabet and Poisson tails [2310.16507].

The DTPC considered there has input alphabet \(X\subseteq \mathbb R_0^+\), output alphabet \(Y=\mathbb Z_0^+\), dark current \(\lambda_0\ge 0\), and transition law
\[
W(y|x)=e^{-(x+\lambda_0)}\frac{(x+\lambda_0)^y}{y!},\qquad y\in \mathbb Z_0^+,
\]
subject to the peak constraint \(0\le x_t\le P_{\max}\) and the average constraint \(\frac1n\sum_{t=1}^n x_t\le P_{\avg}\). Its main theorem states
\[
C_{ID}(W,P_{\max},P_{\avg})=C(W,P_{\max},P_{\avg}),
\]
where
\[
C(W,P_{\max},P_{\avg})=\max_{P_X:0\le X\le P_{\max},\,\mathbb E[X]\le P_{\avg}} I(X;Y).
\]
If \(P_X^*\) is a capacity-achieving input law, then the optimal identification rate is
\[
R_{ID}=C_{ID}=I(P_X^*,W)\;\text{bits/use}.
\]
Thus, although randomized identification codes can support approximately \(2^{2^{nC}}\) hypotheses, the identifying rate measured by \(\frac1n\log\log N\) is limited by exactly the same constant \(C\) as the transmission rate [2310.16507].

The proof structure mirrors the standard identification paradigm. Achievability uses an optimal input distribution \(P_X^*\), subcodes of size approximately \(\exp(nC)\), a decoding rule based on empirical information density relative to the induced output law \(P_Y\), and concentration tools such as Chernoff or Hoeffding bounds. The converse is a strong converse obtained through the information-spectrum method: for every \(\varepsilon>0\),
\[
\Pr\!\left\{\frac1n\log\frac{W^n(Y^n|X^n)}{P_{Y^n}(Y^n)}\ge C+\varepsilon\right\}\to 0,
\]
with Chebyshev’s inequality and moment bounds for Poisson distributions controlling the relevant sums [2310.16507].

The same mechanism extends to a state-dependent DTPC with i.i.d. state \(S^n\sim P_S\), unknown to both parties. Defining the averaged channel
\[
\overline W(y|x)=\sum_s P_S(s)\,W_S(y|x,s),
\]
one obtains
\[
C_{ID}(W_S,P_{\max},P_{\avg})
=\max_{P_X:0\le X\le P_{\max},\,\mathbb E[X]\le P_{\avg}} I_{\overline W}(X;Y)
=C(\overline W,P_{\max},P_{\avg}),
\]
so the identification capacity is again the Shannon capacity of the averaged channel [2310.16507].

## 3. Deterministic identification and dimensional characterizations

Deterministic identification has a markedly different asymptotic structure. For memoryless channels with finite output but arbitrary input alphabet, the channel is represented by its output distributions \(W_x\) as a subset \(\widetilde{\mathcal X}=W(\mathcal X)\subset \mathcal P(\mathcal Y)\). The maximum deterministic code size satisfies
\[
N_{DI}(n,\lambda_1,\lambda_2)=2^{\Theta(n\log n)},
\]
and the corresponding \(n\log n\)-rate is controlled by the Minkowski dimension of \(\widetilde{\mathcal X}\) [2402.09117].

More precisely, if \(\underline d_M\) and \(\overline d_M\) denote the lower and upper Minkowski dimensions, then
\[
\frac14\,\underline d_M(\widetilde{\mathcal X})\le \dot C_{DI}(W)\le \overline d_M(\widetilde{\mathcal X}).
\]
If the ordinary Minkowski dimension
\[
d=\lim_{\delta\to 0}\frac{\log \Gamma_\delta(\widetilde{\mathcal X})}{-\log \delta}
\]
exists, this becomes
\[
\frac d4\le \dot C_{DI}(W)\le d.
\]
The converse uses disjoint total-variation balls of radius \(\delta/n\) around the output distributions corresponding to codewords, while the achievability argument combines packings in \(\widetilde{\mathcal X}\) with classical codes of large Hamming distance [2402.09117].

A key technical ingredient is the “Hypothesis-Testing Lemma.” For codewords \(u,v\in \mathcal X^n\) whose induced output distributions satisfy \(\frac12\|W_u-W_v\|_1\ge 1-\varepsilon\), one chooses as the test region for \(u\) its conditional-entropy typical set
\[
\mathcal T_u^\delta=\{y^n: |-\log W^n(y^n|u)-H(W^n_u)|\le \delta\sqrt n\}.
\]
Classical typicality then yields small first-kind error, while trace-distance separation and equipartition estimates control second-kind error [2402.09117].

The dimensional perspective clarifies why deterministic identification can be super-exponential without becoming doubly exponential. The growth is tied to the metric complexity of the single-letter output set rather than to mutual information. The same detailed exposition states that there is no “superactivation” of deterministic identification capacity for classical channels, because Minkowski dimension under Cartesian product adds and channels with zero single-letter dimension retain zero dimension after product formation [2402.09117]. A plausible implication is that deterministic identification is governed more by output-set geometry than by the combinatorial amplification characteristic of randomized identification.

## 4. Continuous-input classical channels: Poisson and colored Gaussian models

Continuous-input models sharpen the contrast between randomized and deterministic identification. The DTPC with molecule-counting receivers is motivated by event-driven molecular communications, where the conventional Shannon capacity may not be the appropriate metric and identification is proposed as an alternative performance measure. Yet, once randomized encoding is permitted, the DTPC again satisfies \(C_{ID}=C\) under peak and average power constraints [2310.16507].

The colored Gaussian channel with inter-symbol interference presents the deterministic counterpart. The model has \(K=K(n,\kappa)=n^\kappa\) taps, \(\kappa\in[0,\tfrac12)\), Toeplitz convolution matrix \(H^h\), and additive colored Gaussian noise \(Z\sim N(0,\Sigma_n)\) on dimension \(\bar n=n+K-1\). The singular-value spectrum of \(\Sigma_n\) is polynomially bounded as
\[
\sigma_{\min}(\Sigma_n)\ge C_{\min} n^{-\mu},\qquad
\sigma_{\max}(\Sigma_n)\le C_{\max} n^{\mu/2},
\]
with \(\mu\in[0,\tfrac12)\), and the input obeys the per-symbol peak-power constraint \(|x_t|\le P_{\max}\) [2604.04674].

Under assumptions \((A1)\)–\((A3)\) and \(\kappa+\mu<\tfrac12\), the identification capacity \(C_I\) admits super-exponential codebook sizes
\[
|\mathcal C_n|=2^{(n\log n)R+o(n\log n)}
\]
and satisfies
\[
\frac{1-2(\kappa+\mu)}{4}\le C_I\le 1+\kappa+\frac{\mu}{2}.
\]
The achievability proof constructs a packed codebook in \([ -P_{\max},P_{\max}]^n\), passes to the “convoluted codebook” \(c_i^h=H^h c_i\), uses the minimum-distance estimate
\[
\|c_i^h-c_j^h\|\ge H_{\min}\|c_i-c_j\|,\qquad H_{\min}=\inf_\omega |H(\omega)|>0,
\]
and decodes with the Mahalanobis-distance test
\[
T(y,c_j^h)=n^{-1}(y-c_j^h)^T\Sigma_n^{-1}(y-c_j^h)-1,\qquad
D_j=\{|T(y,c_j^h)|\le \delta_n\}.
\]
Type-I and type-II errors vanish by Chebyshev bounds on \(\chi^2\)-type deviations [2604.04674].

In the memoryless white-noise case \((\kappa,\mu)=(0,0)\), one recovers
\[
\frac14\le C_I\le 1,
\]
consistent with earlier deterministic-identification results for the AWGN channel under peak power [2604.04674]. As \(\mu\) increases or \(\kappa\) increases, the achievable region shrinks and vanishes at \(\kappa+\mu=\tfrac12\). This suggests that channel memory and spectral ill-conditioning act as geometric penalties on deterministic identification packings.

## 5. Classical identification over quantum channels

The phrase “classical identification capacity” also refers to the transmission of **classical messages in identification mode through quantum channels**. For a finite-dimensional quantum channel \(N:L(A)\to L(B)\), an \((n,N,\lambda_1,\lambda_2)\) classical identification code consists of code states \(\rho_i\in D(A^{\otimes n})\) and effects \(0\le D_i\le 1_{B^{\otimes n}}\) such that
\[
\operatorname{Tr}[N^{\otimes n}(\rho_i)D_i]\ge 1-\lambda_1,\qquad
\forall j\neq i:\ \operatorname{Tr}[N^{\otimes n}(\rho_i)D_j]\le \lambda_2
\]
[2606.05032].

For stationary memoryless classical–quantum channels \(W:X\to \mathcal S(\mathbb B)\), the established identity is
\[
C_{ID}(W)=C_{ID}^{sim}(W)=C(W)=\max_{P_X}\chi(P_X;W),
\]
and for fixed \(\varepsilon_1+\varepsilon_2<1\) there is a strong converse of the form
\[
N(n,\varepsilon_1,\varepsilon_2)\le \exp\!\big[\exp(n(C(W)+o(1)))\big].
\]
The same equality between identification and transmission capacity extends to compound memoryless cq-channels:
\[
C_{ID}(\mathcal W)=C_{ID}^{sim}(\mathcal W)=C(\mathcal W)=\max_{P_X}\inf_{t\in\Theta}\chi(P_X;W_t),
\]
and, for arbitrarily varying cq-channels, one has the dichotomy
\[
C_{ID}(\{W_t\})=C_{ID}^{sim}(\{W_t\})=
\begin{cases}
0,&\text{if }\{W_t\}\text{ is symmetrizable},\\
C_{ran}(\{W_t\}),&\text{otherwise}.
\end{cases}
\]
These formulas show that the equality between randomized identification capacity and transmission capacity survives substantial channel uncertainty [1801.09967].

Secure identification introduces a second observer. For compound wiretap cqq-channels and arbitrarily varying wiretap cqq-channels, the secret identification capacity obeys a dichotomy: it is \(0\) when the secrecy capacity is \(0\), and otherwise equals the transmission capacity of the main channel. In the compound case,
\[
C_{ID}^S(\mathcal W,\mathcal V)=C_{ID}^{S,sim}(\mathcal W,\mathcal V)=
\begin{cases}
0,&C_S(\mathcal W,\mathcal V)=0,\\
C(\mathcal W),&C_S(\mathcal W,\mathcal V)>0.
\end{cases}
\]
This identifies a threshold phenomenon absent from ordinary randomized identification without secrecy constraints [1801.09967].

## 6. Strong converses and converse geometry

Strong converse bounds for classical identification over quantum channels have become a distinct line of work. For the qubit depolarizing channel
\[
\mathcal N_p(\rho)=(1-p)\rho+\frac p2 I_2,
\]
a strong-converse upper bound on the unrestricted identification capacity is
\[
C_{ID}(\mathcal N_p)\le
\begin{cases}
2,&0\le p\le 1-2^{-2/3},\\
2-D(\gamma(p)\|3/4),&1-2^{-2/3}\le p<1,
\end{cases}
\]
where
\[
\gamma(p)=-\frac{1}{2\log(1-p)},\qquad
D(x\|y)=x\log(x/y)+(1-x)\log[(1-x)/(1-y)].
\]
As \(p\to 1\), this bound tends to \(0\), matching the completely noisy limit [2603.29987].

Under simultaneous identification with complete product measurements, the same channel admits an exact capacity formula:
\[
\tilde C_{ID}^{sim}(\mathcal N_p)=1-h(p/2),
\]
and a strong converse holds at that rate. The proof reduces product-basis measurements on \(\mathcal N_p^{\otimes n}\) to an \(n\)-fold binary symmetric channel with crossover probability \(p/2\), then uses classical soft-covering to control the number of distinguishable output distributions [2603.29987].

A more general converse framework is given by the Gaussian mean-width method. Fixing a full-rank state \(\sigma\in D_+(B)\), one defines the weighted inner product
\[
\langle Y,Z\rangle_\sigma=\operatorname{Tr}[Y\,\sigma^{-1/2}Z\,\sigma^{-1/2}],
\]
the weighted adjoint \(N^{*,\sigma}\), and the single-letter positive operator
\[
\Omega_\sigma=\mathbb E_{G\sim \mathcal N(0,)}\big[(N^{*,\sigma}(G))^2\big].
\]
Combining the domination \(\|Y\|_1\le \|Y\|_\sigma\), Gaussian mean-width estimates in the product geometry, and Sudakov’s inequality yields the single-letter strong converse
\[
C_{ID}(N)\le \inf_{\sigma\in D_+(B)}\log\|\Omega_\sigma\|_\infty.
\]
This bound admits a semidefinite-program representation and improves previously known converse bounds for several channels, including depolarizing, Pauli, erasure, and amplitude damping channels [2606.05032].

One recurring misconception is that identification “exceeds” transmission capacity because it permits vastly more messages. The classical theory distinguishes sharply between **message count** and **rate definition**. Randomized identification indeed supports doubly-exponential hypothesis sets, but its asymptotic rate remains \(C\) for DMCs, the DTPC, and memoryless cq-channels. Deterministic identification, by contrast, replaces the \(\frac1n\log\log N\) scale by \(\frac1{n\log n}\log N\) and is governed by geometric quantities such as Minkowski dimension or packing under Mahalanobis distance [2310.16507] [2402.09117] [2604.04674].

Source: https://www.emergentmind.com/topics/classical-identification-capacity