---
title: Matrix Spencer Conjecture
url: https://www.emergentmind.com/topics/matrix-spencer-conjecture
type: topic
---

# Matrix Spencer Conjecture

Searching arXiv for recent and foundational papers on the Matrix Spencer Conjecture and related special cases.
The Matrix Spencer Conjecture is the matrix analogue of Spencer’s vector-balancing theorem. In its square formulation, it asks whether there is a universal constant \(C>0\) such that for every \(n\) and every family of symmetric or Hermitian contractions \(A_1,\dots,A_n\), one can choose signs \(x\in\{\pm1\}^n\) with \(\bigl\|\sum_{i=1}^n x_iA_i\bigr\|\le C\sqrt n\) [1907.02145]. A more general operator-norm formulation, standard in later work, considers \(A_1,\dots,A_n\in\mathbb R^{m\times m}\) with \(m\ge \sqrt n\) and predicts a bound \(\lesssim \sqrt{\,n\cdot \max(1,\log(m/n))\,}\); in particular, when \(\sqrt n\le m\le n\), the conjectured scale is \(O(\sqrt n)\) [2111.03171]. The problem is motivated by the gap between random-sign matrix concentration, which gives \(O(\sqrt{n\log n})\) in the square case, and the sharper discrepancy scale suggested by the classical diagonal setting [1907.02145].

## 1. Formal statement and asymptotic regimes

A standard formulation is the following: if \(A_1,\dots,A_n\in \mathbb R^{n\times n}\) are symmetric with \(\|A_i\|_{\mathrm{op}}\le 1\), then there should exist \(x\in\{\pm1\}^n\) such that
\[
\Big\|\sum_{i=1}^n x_iA_i\Big\|_{\mathrm{op}}\le C\sqrt n.
\]
Equivalently, if the eigenvalues of each \(A_i\) lie in \([-1,1]\), there is a signing with the spectral norm of the signed sum bounded by \(O(\sqrt n)\). A closely related variation asks for the singular values of the signed sum to be bounded by \(O(\sqrt n)\) [1907.02145].

In the more general \(m\times m\) formulation, with \(m\ge \sqrt n\), the conjecture asserts that for symmetric matrices \(A_1,\dots,A_n\) with \(\|A_i\|_{\mathrm{op}}\le 1\), there exist signs \(x\in\{\pm1\}^n\) such that
\[
\Big\|\sum_{i=1}^n x_iA_i\Big\|_{\mathrm{op}}
\lesssim
\sqrt{\,n\cdot \max(1,\log(m/n))\,}.
\]
The conjecture is open even for \(m=n^{1/2+\varepsilon}\) [2111.03171].

The baseline comparison is given by matrix concentration. A uniform random signing yields with high probability
\[
\Big\|\sum_{i=1}^n x_iA_i\Big\|_{\mathrm{op}}\le C'\sqrt{n\log n},
\]
and the same bound can be obtained deterministically via matrix multiplicative weights. Removing the extra \(\sqrt{\log n}\) factor is the central open step [1907.02145].

The conjecture coincides with classical Spencer-type discrepancy in commuting regimes. For diagonal \(A_i\), it matches Spencer’s “Six Standard Deviations Suffice” theorem up to the matrix-versus-scalar distinction. For commuting matrices, or more generally simultaneously diagonalizable families, vector discrepancy methods apply directly. By contrast, general noncommutative families do not reduce to a finite collection of scalar inequalities [2111.03171].

Several stronger-looking statements are known to fail or to require extra structure. In the Kadison–Singer/Weaver setting, Marcus–Spielman–Srivastava obtain \(O(\varepsilon)\) operator-norm discrepancy for isotropic sums of small rank-one matrices, but that regime assumes \(\|v_i\|_2\le \varepsilon\) and \(\sum v_iv_i^\top=I\), and therefore does not match the general Matrix Spencer normalization [1907.02145]. A variance-sensitive strengthening,
\[
\Bigl\|\sum_{i=1}^n x_iA_i\Bigr\|\le C\Bigl\|\sum_{i=1}^n A_i^2\Bigr\|^{1/2},
\]
has also been proposed, but it is false in full generality: a diagonal-matrix counterexample is proved in the algebraic framework paper [2606.16005].

## 2. Discrepancy bodies and the non-polyhedral obstruction

A central object is the convex body of good fractional signings. Given symmetric matrices \(A_1,\dots,A_m\in\mathbb R^{n\times n}\) and a target bound \(B>0\), define
\[
K(B):=\Bigl\{y\in\mathbb R^m:\Bigl\|\sum_{i=1}^m y_iA_i\Bigr\|_{\mathrm{op}}\le B\Bigr\}.
\]
In the Matrix Spencer regime, one sets \(m=n\), \(\|A_i\|_{\mathrm{op}}\le 1\), and \(B\asymp \sqrt n\); the geometric question is whether \(K(\sqrt n)\) is large enough in Gaussian measure to support discrepancy machinery [1907.02145].

For symmetric matrices,
\[
\|M\|_{\mathrm{op}}=\sup_{u\in S^{n-1}} |u^\top Mu|,
\]
so
\[
K(B)=\bigcap_{u\in S^{n-1}}
\Bigl\{y\in\mathbb R^m:\Bigl|\sum_{i=1}^m y_i\langle A_i,uu^\top\rangle\Bigr|\le B\Bigr\}.
\]
Thus \(K(B)\) is an intersection of infinitely many symmetric strips indexed by the sphere. This is the geometric source of difficulty: classical strip-intersection arguments such as Šidák–Khatri rely on finitely many slabs, whereas the operator-norm body is typically non-polyhedral and has a smooth boundary induced by spectral calculus [1907.02145].

The same issue appears in the discrepancy-body formalism used in mirror-descent approaches. For a norm \(\|\cdot\|\) on \(\mathbb R^{m\times m}\), let
\[
D:=\Bigl\{x\in\mathbb R^n:\Bigl\|\sum_{i=1}^n x_iA_i\Bigr\|\le 1\Bigr\}.
\]
Its polar is
\[
D^\circ
=
\Bigl\{
(\langle A_1,U\rangle,\dots,\langle A_n,U\rangle): \|U\|_*\le 1
\Bigr\},
\]
where \(\|\cdot\|_*\) is the dual norm. A key equivalence established in the mirror-descent framework is that \(\gamma_n(D)\ge 2^{-O(n)}\) if and only if \(D^\circ\) can be covered by \(2^{O(n)}\) translates of \(\frac1n B_\infty^n\); equivalently, \(\gamma_n(tK)\ge 2^{-O(n)}\) if and only if \(N(K^\circ,(t/n)B_\infty^n)\le 2^{O(n)}\) [2111.03171].

This reframes Matrix Spencer as a question about either Gaussian measure or dual covering numbers. The usual random-sign and matrix-concentration arguments show that the desired event is not extremely rare, but they do not supply the precise lower bounds on \(\gamma_n(K)\) needed for constructive partial coloring [1907.02145].

## 3. Gaussian expansion, covering duality, and analytic mechanisms

One major advance concerns a structured non-polyhedral setting arising from spectral sparsification. Under the normalization
\[
\sum_{i=1}^m |A_i| \preceq I_n,
\]
with \(|A|\) defined by replacing eigenvalues by their absolute values in the spectral decomposition, and with \(m=n/\varepsilon^2\), consider
\[
K=\Bigl\{x\in\mathbb R^m:\Bigl\|\sum_{i=1}^m x_iA_i\Bigr\|_{\mathrm{op}}\le \varepsilon\Bigr\}.
\]
The main Gaussian expansion theorem shows that for every \(\alpha\in(0,1)\),
\[
\gamma_m\bigl((50/\alpha)K+\alpha\sqrt m\,B_2^m\bigr)\ge \tfrac12,
\]
and, for any \(\delta>0\),
\[
\gamma_m\bigl((50/\alpha)K+(\alpha+\delta)\sqrt m\,B_2^m\bigr)\ge 1-e^{-\delta^2m/2}.
\]
The same statement remains valid after pinning a small set of coordinates, and the measure guarantees imply the mean-width lower bound \(w(K)=\Omega(\sqrt m)\) [1907.02145].

The proof introduces a barrier-shifted potential
\[
\Phi_{C,D}(x):=\operatorname{tr}(A_{C,D}(x)^{-1}),
\qquad
A_{C,D}(x):=(C+D\|x\|_2^2)I-\sum_{i=1}^m x_iA_i,
\]
and analyzes a Brownian-motion-like process \(x\leftarrow x+\delta y\), where \(y\) is Gaussian but restricted to a large-codimension subspace that removes dangerous directions and preserves linear constraints. The argument relies on a matrix Taylor approximation for the inverse in the noncommuting setting, a trace-of-product inequality
\[
\operatorname{tr}(A_1BA_2B)\le \operatorname{tr}(A_1|B|)\operatorname{tr}(A_2|B|),
\]
and Gaussian isoperimetry. The result is a smooth potential-theoretic substitute for the finite-strip arguments that fail for operator-norm bodies [1907.02145].

A distinct but complementary framework is built around covering the polar body by mirror descent. For the operator norm, the natural mirror map on the spectraplex
\[
\mathcal S_m:=\{X\succeq 0:\operatorname{Tr}X=1\}
\]
is the quantum relative entropy
\[
\Phi(X)=\operatorname{Tr}(X\log X),
\]
which is \(1/2\)-strongly convex with respect to the trace norm. Starting from a set of initial points \(T_0\) with \(|T_0|\le 2^{O(n)}\), one runs mirror descent for \(T=n\) steps against subgradients from \(\{\pm A_1,\dots,\pm A_n\}\). Because each iterate depends only on the initial point and the sum of chosen subgradients, the number of distinct trajectories is at most \(2^{O(n)}\). If the starting set approximates every density matrix in quantum relative entropy by \(D^{\max}\), the resulting centers cover the polar by cubes of radius \(t/n\) with \(t\approx \sqrt{nD^{\max}}\), and hence imply \(\gamma_n(tK)\ge 2^{-O(n)}\). The same framework extends to Schatten norms via the mirror map \(\Phi(X)=\|X\|_{S_p}^2/[2(p-1)]\) for \(p\in(1,2]\) [2111.03171].

These two lines of work address the same bottleneck from different directions. The barrier method proves robust Gaussian expansion for a specific operator-norm body under sparsifier normalization, whereas the mirror-descent framework converts geometric approximation of the dual body into partial coloring. Together they make explicit that Matrix Spencer is fundamentally a measure-and-covering problem for a smooth, non-polyhedral convex body [1907.02145][2111.03171].

## 4. Structured resolutions and quantitative special cases

Several substantial special cases are now known, and they exhibit different mechanisms: low-rank structure, block structure, Frobenius control, and Schatten-norm interpolation.

| Regime | Hypothesis | Guarantee |
|---|---|---|
| Moderate-rank matrices | \(\|A_i\|\le 1\), \(\|A_i\|_F\le n^{1/4}\) | \(\bigl\|\sum x_iA_i\bigr\|\le O(\sqrt{n\log(m/n)})\); randomized polynomial time |
| Low-rank operator norm | \(\|A_i\|_{\mathrm{op}}\le 1\), \(\operatorname{rank}(A_i)\le r\) | \(\bigl\|\sum x_iA_i\bigr\|_{\mathrm{op}}\lesssim \sqrt{n\max(1,\log(r\min\{1,m/n\}))}\) |
| Block-diagonal operator norm | \(\|A_i\|_{\mathrm{op}}\le 1\), common block size \(h\) | \(\bigl\|\sum x_iA_i\bigr\|_{\mathrm{op}}\lesssim \sqrt{n\max(1,\log(hm/n))}\) |
| Poly-logarithmic rank or dimension | \(\|A_i\|_{\mathrm{op}}\le 1\), \(\|A_i\|_F^2\le n/\log^3 n\) | \(\bigl\|\sum x_iA_i\bigr\|_{\mathrm{op}}=O(\sqrt n)\); randomized polynomial time |

The moderate-rank result is derived from a partial-coloring theorem controlled by
\[
A:=\sum_i A_i^2.
\]
It produces \(x\in[-1,1]^n\) with \(\Omega(n)\) integral coordinates and
\[
\Bigl\|\sum_i x_iA_i\Bigr\|
\le
\|A\|^{1/2}\cdot O\!\Bigl(\sqrt{1+\log\!\Bigl(\frac{\operatorname{Tr}(A)}{\sqrt n\,\|A\|}\Bigr)}\Bigr),
\]
together with a randomized polynomial-time algorithm based on semidefinite programming. Under \(\|A_i\|\le 1\) and \(\|A_i\|_F\le n^{1/4}\), this yields \(O(\sqrt{n\log(m/n)})\), and hence \(O(\sqrt n)\) when \(m=\Theta(n)\). The argument is based on a “compress-or-color” dichotomy, one-way quantum communication lower bounds, purification, and sketching [2110.10099].

The mirror-descent framework gives two further operator-norm results. For low-rank matrices, one can efficiently find a coloring with
\[
\Bigl\|\sum_{i=1}^n x_iA_i\Bigr\|_{\mathrm{op}}
\lesssim
\sqrt{\,n\cdot \max\bigl(1,\log(r\cdot \min\{1,m/n\})\bigr)\,}.
\]
When \(rm\le n\), this proves the Matrix Spencer conjecture. For common block-diagonal structure with block size \(h\), one obtains
\[
\Bigl\|\sum_{i=1}^n x_iA_i\Bigr\|_{\mathrm{op}}
\lesssim
\sqrt{\,n\cdot \max(1,\log(hm/n))\,},
\]
removing the earlier assumption \(h\le \sqrt n\). The same paper also proves Schatten-norm partial colorings: for \(2\le p\le q\le \infty\), \(\|A_i\|_{S_p}\le 1\), and \(\operatorname{rank}(A_i)\le r\), there is a polynomial-time algorithm returning \(x\in[-1,1]^n\) with at least \(n/2\) coordinates in \(\{\pm1\}\) and
\[
\Bigl\|\sum_{i=1}^n x_iA_i\Bigr\|_{S_q}
\lesssim
\sqrt{\,n\cdot \min\{p,\max(1,\log(rk))\}\,}\cdot k^{1/p-1/q},
\qquad
k=\min(1,m/n).
\]
A full coloring is then obtained at a controlled loss [2111.03171].

The strongest current low-rank resolution is the poly-logarithmic-rank theorem. If \(\|A_i\|_{\mathrm{op}}\le 1\) and \(\|A_i\|_F^2\le n/\log^3 n\) for all \(i\), then there exist signs \(x\in\{\pm1\}^n\) such that
\[
\Bigl\|\sum_{i=1}^n x_iA_i\Bigr\|_{\mathrm{op}}=O(\sqrt n),
\]
and the signs can be computed in randomized polynomial time. Since \(\|A_i\|_F^2\le \operatorname{rank}(A_i)\|A_i\|_{\mathrm{op}}^2\), this holds in particular when \(\operatorname{rank}(A_i)\le n/\log^3 n\), and also when the ambient dimension \(d\le n/\log^3 n\). The proof combines a refined noncommutative Khintchine inequality for correlated Gaussian entries due to Bandeira–Boedihardjo–van Handel with a partial-coloring lemma of the form
\[
\Bigl\|\sum_{i=1}^n (x_i-x_{0,i})A_i\Bigr\|_{\mathrm{op}}
\le
c\Bigl(\sigma+(\log^{3/4}d)\sqrt{\sigma f}\Bigr),
\]
where \(\sigma^2=\bigl\|\sum_i A_i^2\bigr\|_{\mathrm{op}}\) and \(\sum_i\|A_i\|_F^2\le nf^2\). Iterating for \(O(\log n)\) rounds yields a full coloring. This result also implies a \(\log n-\Omega(\log\log n)\) qubit lower bound for quantum random access codes encoding \(n\) classical bits with advantage \(\gg 1/\sqrt n\) [2208.11286].

## 5. Algebraic and group-theoretic forms

A recent development replaces ambient matrix dimension by intrinsic algebraic complexity. Let \(\mathcal A\) be a finite-dimensional \(C^*\)-algebra with
\[
\dim_{\mathbb C}(\mathcal A)\le C_0n.
\]
If \(A_1,\dots,A_n\in \mathcal A\) are contractions, not necessarily Hermitian, then there exist signs \(x_1,\dots,x_n\in\{\pm1\}\) such that
\[
\Bigl\|\sum_{i=1}^n x_iA_i\Bigr\|\le C(C_0)\sqrt n.
\]
The proof uses the Wedderburn decomposition
\[
\mathcal A \cong \bigoplus_{\alpha=1}^r I_{r_\alpha}\otimes M_{d_\alpha}(\mathbb C),
\qquad
\dim_{\mathbb C}(\mathcal A)=\sum_\alpha d_\alpha^2,
\]
together with a block norm identity stating that multiplicities \(r_\alpha\) do not affect the operator norm of signed sums. The decisive parameter is therefore \(\sum_\alpha d_\alpha^2\), not the ambient dimension. The same paper proves a square-sum block criterion: if a simultaneously block-diagonal family has block sizes \(b_\alpha\) with \(\sum_\alpha b_\alpha^2\le C_1n\), then Matrix Spencer holds with \(O(\sqrt n)\). It also proves stability under low-rank perturbations: if \(A_i=B_i+L_i\), \(B_i\in\mathcal A\), \(\dim_{\mathbb C}(\mathcal A)\le C_0n\), \(\|B_i\|\le1\), \(\|L_i\|\le1\), and \(\|L_i\|_F^2\le n/\log^3(en)\), then there are signs with \(\bigl\|\sum x_iA_i\bigr\|\le C(C_0)\sqrt n\) [2606.16005].

This algebraic theorem resolves the Group Spencer conjecture as a special case. For a finite group \(G\), the left regular representation satisfies
\[
\rho_{\mathrm{reg}}(g)\cong \bigoplus_{\pi\in\widehat G} I_{d_\pi}\otimes \pi(g),
\qquad
\dim_{\mathbb C}(C^*(\rho_{\mathrm{reg}}(G)))=\sum_{\pi\in\widehat G} d_\pi^2=|G|.
\]
Applying the algebraic theorem gives signs \(x_g\in\{\pm1\}\) such that
\[
\Bigl\|\sum_{g\in G} x_g\,\rho_{\mathrm{reg}}(g)\Bigr\|\le C\sqrt{|G|},
\]
and by complete reducibility the same signs control every unitary representation [2606.16005].

A second 2026 paper proves the regular-representation case directly for all finite groups, and in fact yields the same bound for any unitary representation \(\sigma\):
\[
\exists\,\varepsilon\in\{\pm1\}^G
\quad\text{such that}\quad
\Bigl\|\sum_{g\in G}\varepsilon_g\,\sigma(g)\Bigr\|\le C\sqrt{|G|}.
\]
Its proof is Fourier-analytic. Writing
\[
\sum_{g\in G}\varepsilon_g\,\rho(g)
\cong
\bigoplus_{\pi\in\widehat G}
I_{d_\pi}\otimes
\Bigl(\sum_{g\in G}\varepsilon_g\,\pi(g)\Bigr),
\]
one gets the exact identity
\[
\Bigl\|\sum_{g\in G}\varepsilon_g\,\rho(g)\Bigr\|
=
\max_{\pi\in\widehat G}
\Bigl\|\sum_{g\in G}\varepsilon_g\,\pi(g)\Bigr\|.
\]
The argument then combines blockwise Gaussian concentration on irreducible components, Schur orthogonality, noncommutative Khintchine in low dimension, intrinsic freeness inequalities in high dimension, the Gaussian correlation inequality, and a multi-round partial coloring scheme on subsets \(S\subseteq G\) [2606.12181].

## 6. Algorithms, limitations, and open directions

Constructive discrepancy methods remain organized around partial coloring. In the measure-and-covering framework, a lower bound \(\gamma_n(tK)\ge 2^{-O(n)}\) implies, via the Reis–Rothvoss partial-coloring theorem, a polynomial-time procedure returning \(x\in c\cdot tK\cap[-1,1]^n\) with at least \(n/2\) coordinates saturated; iterating \(O(\log n)\) times gives a full coloring at essentially the same scale when the target is \(n^{\Omega(1)}\) [2111.03171]. In the moderate-rank communication-theoretic approach, each round solves the semidefinite program
\[
\text{maximize }\langle g,x\rangle
\quad
\text{subject to }
\Bigl\|\sum_{i=1}^n x_iA_i\Bigr\|\le \Delta,
\quad
x\in[-1,1]^n,
\]
and randomized polynomial time follows by repeating the partial-coloring step over \(O(\log n)\) rounds [2110.10099].

The sparsification work gives a different algorithmic outcome. In the Batson–Spielman–Srivastava setting, the Gaussian expansion theorem is sufficient to run a discrepancy algorithm that samples a linear-size spectral sparsifier using only a logarithmic number of sampling phases. Starting from PSD matrices \(A_i\) with \(\sum A_i=I\), one repeatedly projects a Gaussian sample onto
\[
K\cap [-1,1]^{\operatorname{supp}(s)},
\qquad
K=\Bigl\{x:\Bigl\|\sum x_is_iA_i\Bigr\|_{\mathrm{op}}\le 1000\widetilde\varepsilon\Bigr\},
\]
updates weights by \(s_i\leftarrow s_i(1+x_i)\), and shrinks the support by a constant factor. After \(O(\log m)\) phases one obtains \(|\operatorname{supp}(s)|\le n/\varepsilon^2\) and
\[
(1-O(\varepsilon))I\preceq \sum s_iA_i\preceq (1+O(\varepsilon))I
\]
with probability at least \(1-2^{-\Omega(n)}\) [1907.02145].

Despite these advances, the general conjecture remains open. One major open route is geometric: for
\[
K_{\mathrm{MS}}:=\Bigl\{x\in\mathbb R^n:\Bigl\|\sum_{i=1}^n x_iA_i\Bigr\|_{\mathrm{op}}\le \sqrt n\Bigr\},
\]
prove either \(\gamma_n(K_{\mathrm{MS}})\ge 2^{-cn}\) or at least a Gaussian-expansion statement of the form
\[
\gamma_n(tK_{\mathrm{MS}}+\alpha\sqrt n\,B_2^n)\ge 1-e^{-\Omega(n)}.
\]
A weaker but still significant target is to show \(w(K_{\mathrm{MS}})\ge c\sqrt n\). The barrier-shifted method establishes such statements only under the stronger normalization \(\sum |A_i|\preceq I\) and scaling \(m=n/\varepsilon^2\); it does not currently extend to the full Matrix Spencer normalization \(\|A_i\|_{\mathrm{op}}\le 1\) with \(m=n\) and target \(\sqrt n\) [1907.02145].

A second open route is entropy-theoretic. If one could construct a set \(T\subset \mathcal S_m\) with \(|T|\le 2^{O(n)}\) such that every \(X\in\mathcal S_m\) has a nearby \(Y\in T\) satisfying
\[
D(X\|Y)\lesssim \max\{1,\log(m/n)\},
\]
then the mirror-descent cover would imply the full Matrix Spencer conjecture. Such nets are known in the block-diagonal setting, but not for the full spectraplex [2111.03171].

Further limitations are structural. The \(O(\sqrt n)\) theorem up to poly-logarithmic rank still requires \(\|A_i\|_F^2\le n/\log^3 n\), and removing that condition remains open [2208.11286]. The algebraic theorem requires \(\dim_{\mathbb C}(\mathcal A)\lesssim n\), and its proof is primarily existential rather than a practical algorithm tailored to the underlying entropy nets [2606.16005]. The communication-theoretic program identifies a conjectural one-way quantum lower bound in the small-advantage regime that would imply full Matrix Spencer; that conjecture is proved in the classical case and for the quantum moderate-rank regime, but not in general [2110.10099].

Two common overgeneralizations are now explicitly ruled out. First, the variance-sensitive bound \(\bigl\|\sum x_iA_i\bigr\|\lesssim \bigl\|\sum A_i^2\bigr\|^{1/2}\) fails in general by a diagonal counterexample [2606.16005]. Second, a matrix analogue of the Komlós conjecture is false: there exist symmetric \(A_i\) with \(\|A_i\|_F\le 1\) such that every signing satisfies
\[
\Bigl\|\sum_{i=1}^n x_iA_i\Bigr\|_{\mathrm{op}}\gtrsim \sqrt{\min(m,n)}.
\]
Accordingly, current progress is best understood not as a uniform reduction to variance or Frobenius control, but as a collection of deep structured theorems that isolate when noncommutative discrepancy can match the \(O(\sqrt n)\) scale [2111.03171].

Source: https://www.emergentmind.com/topics/matrix-spencer-conjecture