--- title: Unsourced Random Access (URA) url: https://www.emergentmind.com/topics/unsourced-random-access-ura type: topic --- # Unsourced Random Access (URA) Unsourced random access (URA) is a grant-free multiple-access paradigm in which a very large population of devices share a single common codebook and the receiver is required to recover only the unordered list of transmitted messages, not the sender identities. In the canonical formulation, \(K_a \ll K_T\) active devices transmit over a frame of length \(n\), each selecting one codeword from a codebook of size \(2^B\); the base station outputs a list \(\widehat{\mathcal L}\) and performance is measured primarily by per-user probability of error (PUPE). This formulation turns random access into a coding problem and has developed from AWGN finite-blocklength theory to asynchronous, fading, multi-antenna, cell-free, secure, and sensing-aware variants [2409.14911], [2509.25074]. ## 1. Canonical formulation In the standard URA model, all users share a common codebook \(\mathcal C=\{\mathbf c(1),\ldots,\mathbf c(2^B)\}\subset\mathbb R^n\). If user \(i\) wishes to send message index \(w_i\in\{1,\ldots,2^B\}\), it transmits \(\mathbf x_i=\mathbf c(w_i)\), and the receiver observes the Gaussian multiple-access channel \[ \mathbf y=\sum_{i=1}^{K_a}\mathbf x_i+\mathbf z,\qquad \mathbf z\sim\mathcal N(\mathbf 0,\sigma^2\mathbf I_n). \] The decoder outputs an unordered list \(\widehat{\mathcal L}\subseteq\{1,\ldots,2^B\}\) of size \(|\widehat{\mathcal L}|=K_a\), ideally matching the transmitted set \(\{w_1,\ldots,w_{K_a}\}\). The survey literature uses PUPE, \[ P_e=\frac1{K_a}\sum_{i=1}^{K_a}\Pr[w_i\notin\widehat{\mathcal L}], \] and spectral efficiency \[ \eta=\frac{K_aB}{n} \] as the canonical performance descriptors [2409.14911]. This “unsourced” formulation removes identity recovery from the physical layer. The monograph literature states that the receiver need only recover the multiset of transmitted messages up to permutation, while the common-codebook assumption collapses the identity problem and avoids explicit ID embedding [2509.25074]. In practice, this has made URA particularly attractive for sporadic short-packet mMTC, where grant-based procedures incur significant delay and signaling overhead [2409.14911]. Many practical URA designs reduce codebook dimensionality by splitting the payload across sections, slots, or sub-slots. In SPARC-based schemes, \(B=LJ\) and one section index is chosen per sub-slot; in uncoupled UURA, \(B=S\cdot J\) and each sub-message is mapped to a shared codebook column in a corresponding sub-slot [1901.06234], [2405.03191]. The same abstraction extends beyond AWGN. For quasi-static Rayleigh fading with \(M\) receive antennas, a common model is \[ Y=\sum_{k=1}^{K_a}x_k h_k^T+Z=XH^T+Z, \] with \(x_k\in\mathbb C^n\), \(h_k\in\mathbb C^M\), and \(Z\) AWGN [2304.12058]. ## 2. Error metrics and information-theoretic limits URA theory is organized around finite-blocklength achievability and converse bounds, energy-per-bit thresholds, and asymptotic many-user limits. The monograph literature summarizes a multi-user converse based on list-decoding and Fano-type arguments, and a random-coding achievability bound of the form \[ P_e \le p_0+\sum_{t=1}^{K_a}\frac{t}{K_a}p_t, \] where \(p_0\) captures collision or power-violation events and each \(p_t\) is exponentially bounded through Gallager-style arguments [2509.25074]. The survey likewise emphasizes non-asymptotic converse and achievability results for GMAC, single-antenna fading, and MIMO fading, with the operative design objective being minimum \(E_b/N_0\) under a PUPE constraint [2409.14911]. A central asymptotic result is that SPARC-based concatenated constructions attain the symmetric Shannon MAC capacity in the unsourced setting. Both “SPARCs for Unsourced Random Access” and “Unsourced Multiuser Sparse Regression Codes achieve the Symmetric MAC Capacity” show vanishing PUPE provided \[ K_aR<\frac12\log_2(1+K_a\mathrm{SNR}), \] thereby lifting point-to-point SPARC optimality to the unsourced many-user regime [1901.06234], [2001.04217]. In these constructions, an inner SPARC stage creates an effective OR-MAC, and an outer code resolves section collisions. The information-theoretic MAP threshold and the algorithmic AMP threshold are not identical. The SPARC capacity paper derives an AMP-success condition for the inner decoding that is strictly below the Shannon limit, while also noting that spatial coupling or power allocation can improve the algorithmic region [2001.04217]. The earlier SPARC paper further gives a linear-program-based power-allocation optimization for the inner code and reports an AMP channel-strength gain of \(\sim 0.5\)–\(1\) dB in regimes where uniform-power AMP has a gap to the MAP threshold [1901.06234]. This separation between optimal decoding and low-complexity decoding remains one of the defining structural tensions in URA. ## 3. Core coding and inference architectures A dominant practical architecture is coded compressed sensing (CCS) and its SPARC variants. The common pattern is an inner sparse superposition or compressed-sensing stage that produces section-wise candidate lists, followed by an outer tree code, random cover code, or OR-MAC code that stitches the sections into valid messages [1901.06234]. This divide-and-conquer strategy is computationally tractable and remains a baseline throughout the literature [2409.14911]. URA has also been generalized to demixing settings. In two-class multi-class URA, a joint AMP+BP decoder over class-specific sparse representations and LDPC-based outer structure yields up to \(\approx 2\) dB gain over SIC at PUPE \(=0.05\) for each class, while treating multi-class access as a coded demixing problem [2102.07704]. The more general coded-demixing framework extends CCS to multiple sparse domains and reports \(\approx 0.5\)–\(0.6\) dB gains over treat-interference-as-noise and SIC in a two-group setting, plus \(\approx 0.2\)–\(0.3\) dB over standard CCS-AMP when a single class is stochastically binned into \(G=2\) subdomains [2203.00239]. Random-spreading URA forms another important family. In the iterative Gaussian-approximation decoder for RS-URA, users select spreading signatures from a common codebook and transmit encoded BPSK payloads over the resulting outer-product structure; the receiver iterates an elementary signal estimator and soft decoders using extrinsic LLR exchange. The paper reports that the decoder stays within \(1\) dB of the theoretical GMAC bounds even at \(K_a=200\), saves \(0.1\)–\(0.2\) dB over the RS-Polar baseline for \(K_a\le 175\), and widens to nearly \(4\) dB at \(K_a=250\) [2512.17628]. Tensor-based URA addresses user separation through tensor decomposition rather than explicit stitching. In “Unsourced Random Access With Tensor-Based and Coherent Modulations,” TBM is combined with a coherent NOMA sub-block; the resulting TBMC scheme outperforms pure TBM over all \(K_a\), approaches the achievability bound for \(K_a<40\), is better than FASURA at low \(K_a\), and is more robust to large-scale fading variance [2304.12058]. “Polar-Coded Tensor-Based Unsourced Random Access with Soft Decoding” adds soft Grassmannian demodulation and an iterative Bayesian receiver with feedback, reporting a \(1\)–\(2\) dB gain at PUPE \(=0.1\) for \(K_a>800\) over BCH-coded BTURA, a \(0.2\) dB gain over FASURA at \(K_a=500\) growing to \(3.6\) dB at \(K_a=1100\), and measured run-time \(30\)–\(50\%\) lower than FASURA in typical settings [2406.16381]. ## 4. Fully asynchronous URA and preamble-free starting-time detection Fully asynchronous URA removes frame alignment altogether. In the ODMA-based model of “A Fully Asynchronous Unsourced Random Access Scheme,” user \(i\) starts at integer time \(d_i\in[0,T]\), and the receiver observes \[ y_j=\sum_{i:\,j\in[d_i,d_i+n-1]}x_i[j-d_i]+z_j,\qquad z_j\sim N(0,\sigma^2). \] Each user splits its \(B\)-bit payload into a \(B_p\)-bit pattern-selection field and \(B_c=B-B_p\) coded data bits, selects one column of a common pattern matrix \(P\in\{0,1\}^{n\times M_p}\) with Hamming weight \(n_a\), polar-encodes the data, and places the \(n_a\) coded symbols on the “on” positions of the selected sparse pattern [2504.11131]. The receiver uses a double sliding-window decoder. The outer window spans \(N_s\times n\) channel uses; the inner window has fixed length \(2n\), which the paper states is the minimum \(\ge n\) needed to resolve pattern overlap. Inside each inner window, the receiver iteratively performs preamble-free joint starting-time and pattern detection, single-user polar-SCL decoding with CRC, and SIC. Starting-time detection is based on the pattern-matched energy \[ e_{k,b}=\|y^{(b)}\circ p_k\|_1, \] computed over candidate start positions \(b=1,\ldots,n\) and pattern indices \(k=1,\ldots,M_p\); the strongest \((\lambda n+u)\) hypotheses are retained rather than thresholding against a fixed preamble detector [2504.11131]. A defining feature of this design is that the sparse ODMA patterns themselves act as “signatures,” eliminating the need for a preamble for starting-time detection. The numerical results are explicit: inner-window length \(2n\) outperforms \(3n\) and \(4n\) by up to \(1.2\) dB at \(K_a=75\); preamble-free ODMA yields \(\sim 1.5\) dB gain over a preamble-based variant; the full asynchronous scheme gains up to \(5.5\) dB over the prior fully asynchronous URA scheme for \(\lambda n\le 100\); and a single-window ODMA design is \(\sim 3\) dB worse than the double-window design [2504.11131]. The paper also identifies a detection error floor at high load, with degradation for \(\lambda n>100\) due to pattern-collision overload, and lists pattern design, adaptive window sizing, MIMO extension, and interference-variance estimation as future work. ## 5. Fading, MIMO, and distributed-network variants In fading and multi-antenna settings, URA departs sharply from the single-antenna GMAC picture. A foundational massive-MIMO result is that covariance-based activity detection can break the classical \(K_a