--- title: Refined Trace–Determinant Inequality url: https://www.emergentmind.com/topics/refined-trace-determinant-inequality type: topic --- # Refined Trace–Determinant Inequality A refined trace–determinant inequality is best understood as an inequality in which a determinant is controlled not merely by total trace, but by additional spectral, block, or geometric data. In the most elementary positive-semidefinite setting, the baseline comparison is the trace-only bound \(\det A\le (\operatorname{tr}A/n)^n\). Refinements replace this coarse control by sharper invariants such as \(\operatorname{tr}(A^2)\), partial traces, off-diagonal products, complete symmetric polynomials, or four-point order data. Current literature does not use one uniform formalism for all such results, but it exhibits a coherent family of determinant bounds that are stronger than trace-only estimates and are often naturally read as trace–determinant statements in an extended sense [1909.05420] [2003.00904] [2003.04520] [2008.05469] [2007.13155]. ## 1. Conceptual structure For positive semidefinite matrices, the determinant is the product of eigenvalues and the trace is their sum. A crude trace bound therefore arises from AM–GM, while a refined trace–determinant inequality supplements \(\sum_i \lambda_i\) by further information that constrains \(\prod_i\lambda_i\) more tightly. In the recent literature, the extra information takes several recurrent forms: a fixed second spectral moment \(\sum_i\lambda_i^2\), blockwise partial traces \(\operatorname{tr}_1\) and \(\operatorname{tr}_2\), higher symmetric polynomials \(s_r\), off-diagonal cycle products, or matrix-order configurations such as \(A\le B\le C\) with \(D=A+C-B\) [1909.05420] [2003.04520] [2008.05469]. This suggests an umbrella notion—*Editor's term*: “refined trace–determinant inequality”—for results that sharpen determinant control by retaining more than the total trace. Under this view, some papers provide direct determinant bounds, while others provide trace inequalities that become determinant statements after exponentiation, specialization to commuting matrices, or use of \(\log\det A=\operatorname{Tr}\log A\). A recurrent misconception is that all such refinements are of the same type. They are not: some are purely algebraic, some are majorization-theoretic, some are block-matrix inequalities, and some are trace-exponential inequalities whose determinant content is indirect [2007.13155] [1011.6325] [1708.04836]. ## 2. Correlation matrices and second-moment refinement A particularly clean model is furnished by correlation matrices. If \(R=(r_{ij})\) is an \(n\times n\) correlation matrix, then \(\operatorname{tr}(R)=n\), so determinant inequalities are automatically trace-constrained. Olkin’s classical comparison used the average off-diagonal correlation \[ r_1=\frac{1}{n(n-1)}\sum_{i\ne j} r_{ij} \] and yielded \[ \det R \le (1-r_1)^{n-1}\bigl(1+(n-1)r_1\bigr). \] The 2019 refinement replaces \(r_1\) by the root-mean-square \[ r_2=\left(\frac{1}{n(n-1)}\sum_{i\ne j} r_{ij}^2\right)^{1/2}, \] and proves the sharper upper estimate \[ \det R \le (1-r_2)^{n-1}\bigl(1+(n-1)r_2\bigr), \] together with the lower comparison \[ (1+r_2)^{n-1}\bigl(1-(n-1)r_2\bigr)\le \det R. \] The mechanism is spectral: \(R\), the equicorrelation matrix with off-diagonal \(+r_2\), and the equicorrelation matrix with off-diagonal \(-r_2\) have the same trace and the same second spectral moment, and their eigenvalue lists form a variance-majorization sandwich [1909.05420]. The determinant consequence is naturally a refined trace–second-moment–determinant statement. Writing the eigenvalues of \(R\) as \(\lambda_1,\dots,\lambda_n\), one has \[ \sum_{i=1}^n\lambda_i=n,\qquad \sum_{i=1}^n\lambda_i^2=n+n(n-1)r_2^2, \] and the theorem says that among correlation-matrix spectra with these first two moments, the determinant is controlled by the equicorrelation spectrum \[ \{1-r_2,\dots,1-r_2,1+(n-1)r_2\}. \] The refinement is genuinely stronger than Olkin’s whenever \(r_1\ge 0\), because the scalar function \(f(t)=(1-t)^{n-1}(1+(n-1)t)\) is decreasing on \([0,1]\) and \(r_2\ge r_1\). The paper also shows that this dominance is not uniform for \(r_1<0\); in that regime the \(r_2\)-bound is an extension, but not always a pointwise improvement [1909.05420]. ## 3. Partial traces, block matrices, and tensor-power hierarchies For positive semidefinite block matrices \(H=[H_{ij}]_{i,j=1}^n\in M_n(M_k)\), the natural trace data are the two partial traces \[ \operatorname{tr}_1 H=\sum_{i=1}^n H_{ii}\in M_k,\qquad \operatorname{tr}_2 H=[\operatorname{tr}H_{ij}]_{i,j=1}^n\in M_n. \] A basic refined determinant inequality in this setting is the strengthened Fiedler–Markham estimate \[ \left(\frac{\det(\operatorname{tr}_2 H)}{n^n}\right)^k\ge \det H, \] equivalently \[ \det(\operatorname{tr}_2 H)^k\ge n^{nk}\det H. \] By permutation similarity between the two block directions, this is equivalent to \[ \left(\frac{\det(\operatorname{tr}_1 H)}{k^k}\right)^n\ge \det H. \] The same block-transposition symmetry also yields Choi-type companion inequalities such as \[ \left(\frac{\operatorname{tr}(\det_1 H)}{k}\right)^k\ge \det H, \] where \(\det_1\) is the first partial determinant. The underlying proof strategy combines permutation similarity, positivity of diagonal blocks, Fischer’s inequality, and Fan–Ky log-concavity [2001.02345]. This block-matrix viewpoint admits a systematic tensor-power lifting. For \(r\in\mathbb N\), one forms block matrices from \(\otimes^r H_{ij}\) or \(\vee^r H_{ij}\), proves positivity of the lifted block matrix, and then applies determinant inequalities there. This produces two trace-side hierarchies: \[ \det\bigl([(\operatorname{tr}H_{ij})^r]_{i,j=1}^n\bigr)\ge k^n(\det H)^{r/k}, \] and \[ \det\bigl([s_r(H_{ij})]_{i,j=1}^n\bigr)\ge \binom{k+r-1}{r}^{\,n}(\det H)^{r/k}, \] where \(s_r\) is the \(r\)-th complete symmetric polynomial of the eigenvalues. A parallel determinant-of-blocks hierarchy is \[ \det\bigl([(\det H_{ij})^r]_{i,j=1}^n\bigr)\ge (\det H)^r. \] These are extensions of the Fiedler–Markham and Thompson inequalities to higher symmetric/tensor powers, and they are explicitly framed as refined determinant inequalities for positive semidefinite block matrices [2003.00904]. A separate line of work strengthens partial-trace determinant inequalities themselves. For \(H\ge 0\in M_n(M_k)\), \[ \det(\operatorname{tr}_2 H)^{kn}\ge k^{kn}\det H,\qquad \det(\operatorname{tr}_1 H)^{nk}\ge n^{nk}\det H, \] and if the numerical range satisfies \(W(H)\subset S_\alpha\), then \[ \det(\operatorname{tr}_1 H)^{nk}\ge (n\cos\alpha)^{nk}|\det H|. \] These are combined with mixed trace–determinant inequalities such as \[ (\operatorname{tr}A)^{mn}+\det(\operatorname{tr}_1 A)^m \ge m^{mn}\bigl(\det A+\det(\operatorname{tr}_2 A)^n\bigr), \] and its \(\operatorname{tr}_1/\operatorname{tr}_2\)-swapped analogue. The same paper extends these statements from positive semidefinite matrices to matrices with numerical range in a sector [2003.04520]. An especially sharp two-sided refinement is the absolute-value inequality \[ (\operatorname{tr}A)^{mn}-\det(\operatorname{tr}_2A)^n \ge \left|\det A-\det(\operatorname{tr}_1A)^m\right| \] for \(A\in M_m(M_n)\) positive semidefinite. This improves Lin’s earlier one-sided inequality by showing that the scalar-trace defect controls the discrepancy between \(\det A\) and the determinant of the first partial trace in absolute value, not only in one sign [2002.09652]. ## 4. Four-point trace inequalities and determinant isoperimetry A different refinement paradigm replaces moment data by order structure. For self-adjoint matrices with spectrum in an interval \((a,b)\), a function \(f\) is called trace minmax if \[ \operatorname{tr}f(A)+\operatorname{tr}f(C)\ge \operatorname{tr}f(B)+\operatorname{tr}f(D), \qquad D=A+C-B, \] whenever \(A\le B\le C\). The exact classification is that \(f\) is trace minmax if and only if \(f'\) is matrix monotone on \((a,b)\), equivalently if and only if \(f'\) analytically continues to a self-map of the upper half-plane. The same theorem gives an integral representation \[ f(z)=\alpha+\beta z+\int \frac{-\log(1-t(z-c))-t(z-c)}{t^2}\,d\mu(t), \] with the interval of integration determined by \((a,b)\) and the center \(c\) [2008.05469]. Exponentiation then turns the trace inequality into a determinant inequality. Writing \(g=e^{-f}\), the determinant-isoperimetric condition is \[ \det g(A)\det g(C)\le \det g(B)\det g(D), \qquad D=A+C-B, \] and the paper proves the equivalence \[ f\ \text{trace minmax} \iff e^{-f}\ \text{determinant isoperimetric}. \] This yields concrete determinant inequalities from matrix-monotone derivatives. For \(f(x)=-\log x\) on \((0,\infty)\), one obtains \[ \det A\,\det C\le \det B\,\det D. \] For \(f(x)=x^2\), one gets \[ \det e^{B^2}\det e^{D^2}\le \det e^{A^2}\det e^{C^2}. \] Affine \(f\) give equality. The same paper also proves that trace minmaxity implies matrix convexity. A common misconception is that ordinary convexity suffices for the four-point determinant inequality; the exact criterion is stronger and is governed by matrix monotonicity of \(f'\), not by scalar convexity alone [2008.05469]. ## 5. Off-diagonal corrections and trace of determinant operators Another family of refinements strengthens Hadamard’s inequality by inserting explicit off-diagonal correction terms. For a positive semidefinite matrix \(A=(a_{ij})\) and any non-identity permutation \(\sigma\in S_n\), \[ \det(A)+\prod_{i=1}^n |a_{i,\sigma(i)}| \le \prod_{i=1}^n a_{ii}. \] When \(\sigma\) is a derangement, this is first proved in that form and then extended to arbitrary nontrivial permutations. Combining with AM–GM on the diagonal entries gives the trace-refined consequence \[ \det(A)\le \left(\frac{\operatorname{tr}A}{n}\right)^n-\prod_{i=1}^n |a_{i,\sigma(i)}|. \] In the block setting, Thompson’s theorem yields \[ \det A+\prod_{i=1}^n |\det A_{i,\tau(i)}| \le \prod_{i=1}^n \det A_{ii} \] for block positive definite matrices and derangements \(\tau\). This line of work is not organized around \(\operatorname{tr}(A^2)\) or partial traces; instead it refines determinant control by subtracting a concrete nonnegative off-diagonal contribution before any passage to trace bounds [2007.13155]. A more noncommutative determinant-trace framework appears for commuting \(d\)-tuples \(\boldsymbol T=(T_1,\dots,T_d)\). The paper defines a symmetrized determinant \(dEt\big([\![\boldsymbol T^*,\boldsymbol T]\!]\big)\) of the block commutator matrix \[ [\![\boldsymbol T^*,\boldsymbol T]\!] = \big([T_j^*,T_i]\big)_{i,j=1}^d, \] and proves that this operator equals the Helton–Howe generalized commutator of the \(2d\)-tuple \((T_1,T_1^*,\dots,T_d,T_d^*)\). If \(\boldsymbol T\) is \(d\)-normal, then this determinant vanishes by the Amitsur–Levitzki theorem. Under the positivity and compression-growth hypotheses defining the class \(BS_{m,\vartheta}(\Omega)\), the determinant is trace class and satisfies \[ \operatorname{trace}\big(dEt([\![\boldsymbol T^*,\boldsymbol T]\!])\big) \le m\,\vartheta\, d!\,\prod_{i=1}^d\|T_i\|^2. \] For the Hardy polydisc coordinate tuple, the paper shows \[ \operatorname{trace}\big(dEt([\![M^*,M]\!])\big)=d!, \] so the bound is sharp in that model. It further conjectures the geometric refinement \[ \operatorname{trace}\big(dEt([\![\boldsymbol T^*,\boldsymbol T]\!])\big) \le \frac{m\,d!}{\pi^d}\,v(\Omega), \] which would be a direct multivariable analogue of Berger–Shaw type spectral-area estimates [2012.11115]. ## 6. Operator-theoretic, quantum, and infinite-dimensional extensions Several papers broaden the trace–determinant theme without always making the determinant statement primary. One example is the inequality \[ \operatorname{Tr}\left((A-B)(B^{-1}-A^{-1})+(C-D)\big((B+D)^{-1}-(A+C)^{-1}\big)\right)\ge 0 \] for \(A,B\succ 0\) and \(C,D\succeq 0\). Its direct subject is a trace inequality with inverse monotonicity, but the standard differential identity \[ \frac{d}{dt}\log\det(X+tH)=\operatorname{Tr}\big((X+tH)^{-1}H\big) \] makes it naturally suggestive for log-determinant comparisons. Similarly, for nonnegative operator monotone \(f\) on \([0,\infty)\) with \(f(0)=0\), \[ \operatorname{Tr}\big((A-B)(f(A)-f(B))\big)\le \operatorname{Tr}\big(|A-B|\,f(|A-B|)\big), \] while the inequality reverses for operator convex \(f\). The case \(f(t)=\log(1+t)\) is especially close to determinant applications because \(\log\det(I+X)=\operatorname{Tr}\log(I+X)\), although these papers do not themselves state universal determinant theorems [1011.6325] [1904.01961]. A second direction is multivariate trace-exponential analysis. The multivariate Golden–Thompson and Araki–Lieb–Thirring inequalities bound quantities such as \[ \log \bigl\|\exp(\sum_{k=1}^n H_k)\bigr\|_p \] by explicit \(\beta_0(t)\)-averages of rotated products, and in the three-matrix case this strengthens Lieb’s triple matrix inequality. A later reformulation replaces complex matrix powers by resolvents, Fréchet derivatives, tensor-product embeddings, and maximally entangled state projectors, thereby converting the multivariate trace inequality into a resolvent form that is more suitable for perturbation theory. These are not determinant inequalities as stated, but they are determinant-compatible because \(\log\det A=\operatorname{Tr}\log A\) in the positive-definite setting, and the four-matrix logarithmic consequence \[ d\,\exp\!\left(\frac1d\operatorname{Tr}\sum_{k=1}^4 \log A_k\right)\le \operatorname{Tr}[P_1\,T_X(Y)] \] is explicitly determinant-like on its left-hand side [1604.03023] [1708.04836]. Recent quantum work sharpens logarithmic trace control itself. For positive semidefinite \(\rho,\sigma\) with \(\operatorname{supp}(\rho)\subseteq \operatorname{supp}(\sigma)\), one paper proves the optimal inequality \[ \operatorname{Tr}\!\big[\rho(\log(\rho+\sigma)-\log\sigma)\big]\le G_s\,Q_{1+s}(\rho\|\sigma), \qquad 0