Shifted Partial Derivative Dimension

• Shifted Partial Derivative Dimension is an algebraic measure that quantifies the linear-algebraic richness of polynomials through shifted partial derivative spaces.
• It employs the SPDP matrix framework to derive dual invariants—SPDP rank and codimension—that underpin circuit lower bound arguments.
• Its applications span separating computational complexity classes and refining methodologies for upper bound analysis in algebraic complexity.

The shifted partial derivative dimension (SPD-dimension) is a central algebraic complexity measure for polynomials, quantifying the size of the space spanned by suitably shifted partial derivatives. This measure, crucial in circuit lower bound arguments, captures the linear-algebraic richness of a polynomial under prescribed differentiation and shift operations. The explicit matrix formalism introduced in the Shifted Partial Derivative Polynomial (SPDP) framework canonically encodes this space, yielding dual measures: SPDP rank, which is the SPD-dimension itself, and SPDP codimension, quantifying the deficit from ambient fullness. This dimension-theoretic approach underpins key arguments in the separation of polynomials such as the permanent and determinant, and it enables both upper and lower bounds on circuit complexity (Edwards, 23 Dec 2025, Efremenko et al., 2016).

1. Formal Definitions and Matrix Formalism

Let $p \in F[x_1,\dots,x_n]$ be a polynomial of degree $d$ over a field $F$. For nonnegative integers $k$ (order of differentiation) and $\ell$ (shift bound), define the classical SPD space as

$$\mathsf{SPD}_{k,\ell}(p) := \mathrm{span}_F\left\{ m \cdot \partial^\alpha p : |\alpha| = k, \deg(m) \leq \ell \right\}$$

where $$\partial^\alpha = \partial_{x_1}^{\alpha_1}\cdots\partial_{x_n}^{\alpha_n}$$. The SPD-dimension is $\SPDdim_{k,\ell}(p) = \dim_F(\mathsf{SPD}_{k,\ell}(p))$.

The SPDP framework refines this by specifying the generating family: $\Gcal_{k,\ell}(p) := \{ m \cdot \partial_S p : S \subseteq [n], |S| = k, m \in \Mmon_{\leq \ell} \}$ where $\partial_S p$ denotes the product of partial derivatives indexed by $S$ and $\Mmon_{\leq \ell}$ is the set of monomials of degree at most $\ell$. The linear span $V_{k,\ell}(p) := \mathrm{span}_F(\Gcal_{k,\ell}(p))$ coincides with $\mathsf{SPD}_{k,\ell}(p)$.

To make codimension meaningful, the SPDP matrix $M_{k,\ell}(p)$ is formed with rows indexed by generator pairs $(S, m)$ and columns by the ambient basis $B_{k,\ell}$ of monomials of degree at most $D = \max\{0, d - k + \ell\}$. Each row represents the coefficient vector of $m \cdot \partial_S p$.

Two dual invariants:

• SPDP rank: $$\Gamma_{k,\ell}(p) := \mathrm{rank}(M_{k,\ell}(p)) = \dim_F V_{k,\ell}(p)$$
• SPDP codimension: $\codim_{k,\ell}(p) := N_{k,\ell}(p) - \Gamma_{k,\ell}(p)$, where $N_{k,\ell}(p) = |B_{k,\ell}|$

Codimension precisely measures the gap between the shifted-derivative span and the ambient space (Edwards, 23 Dec 2025).

2. Linear-Algebraic Structure and Invariances

The SPDP matrix framework gives a concrete linear-algebraic realization of SPD-dimension:

• The row span of $M_{k,\ell}(p)$ equals $V_{k,\ell}(p)$, so rank and SPD-dimension coincide.
• Invariance under basis change: choosing a different ambient basis for $W_{k,\ell}(p)$ changes $M_{k,\ell}(p)$ by right multiplication with an invertible matrix, preserving rank and codimension.
• Variable permutation invariance: permuting variables acts by simultaneous row and column permutations, which also leave rank and codimension unchanged.
• Robustness: All definitions and invariances persist under multilinear/Boolean quotient embeddings by working modulo $(x_i^2 - x_i)$, specializing the ambient basis to the set of square-free monomials (Edwards, 23 Dec 2025).

3. Monotonicity and Structural Properties

Several monotonicity results and structural properties hold:

• Monotonicity in shift ($\ell$): $\ell' \geq \ell$ implies $\Gamma_{k,\ell}(p) \leq \Gamma_{k,\ell'}(p)$ and $\codim_{k,\ell}(p) \geq \codim_{k,\ell'}(p)$.
• Monotonicity in derivative order ($k$): Allowing all derivatives up to order $k$ gives a nondecreasing sequence for cumulative rank, i.e. $$\Gamma_{\leq k,\ell}(p) \leq \Gamma_{\leq k+1,\ell}(p)$$.
• Codimension viewpoint: Codimension directly quantifies the failure of shifted partials to span the ambient space, which is often the crux in lower-bound arguments (Edwards, 23 Dec 2025).

4. Upper Bound Methodologies and Model-Agnostic Templates

A central generic upper-bound template is the “width→rank” schema, which abstracts the influence of local width restrictions in circuit models on SPD-dimension:

• For computations with interface ports in blocks, local operations impact only a bounded number of interfaces in each derivative step.
• The complexity is reflected in the histogram of local “profiles,” leading to at most polynomially many possibilities, each contributing a low-dimensional subspace.
• Summing over profiles enables a model-agnostic upper bound: $$\Gamma_{k,\ell}(p) \leq \sum_h \dim V_h = (\log n)^{O(1)} \leq n^{O(1)}$$ This methodology remains independent of the detailed structure of the circuit, with model-specific refinements and optimizations separated from the core SPDP formalism (Edwards, 23 Dec 2025).

5. Illustrative Examples and Lower Bound Applications

Toy Example: For $p(x_1, x_2, x_3) = x_1x_2 + x_2x_3$, $(k,\ell) = (1,1)$, the SPDP matrix $M_{1,1}(p)$ has dimensions $12 \times 7$ (rows: 3 choices of derivative variable, 4 choices of monomial shift; columns: 7 basis monomials in degree up to 2). The observed rank is 4, giving codimension 3 (Edwards, 23 Dec 2025).

Permanent rank lower bound: For the $n \times n$ permanent, $k = \lfloor n/2 \rfloor, \ell = 0$, the space of shifted partials has dimension at least $\binom{n}{k}$, which is exponential for large $n$, while the codimension is the gap to ambient fullness (Edwards, 23 Dec 2025).

Comparisons for permanent vs. determinant: The SPD-dimension has been pivotal in demonstrating that, for large $n$, the shifted partials of the padded permanent polynomial remain strictly lower-dimensional than those of the determinant, under all relevant parameter regimes. Three forms of degeneration (block-diagonal, two-term complete intersection, initial monomial flattening) realize tight upper bounds on the SPD-dimension of determinant degenerations, cementing the approach's limitations for distinguishing permanent from determinant in orbit closure problems (Efremenko et al., 2016).

6. Limitations and Perspectives

The SPD-dimension, while powerful for depth-4 circuit lower bounds, has known limitations. For example, arguments based purely on SPD-dimension can at best show a quadratic separation between the permanent and determinant, and cannot yield super-polynomial lower bounds for this comparison when $n > 2m^2 + 2m$ (Efremenko et al., 2016). This suggests the need for more refined invariants, such as Young flattenings or syzygy-based measures, to achieve stronger separations. The codimension perspective in the SPDP framework systematizes the identification of failing cases and clarifies the intrinsic limitation of shifted-partial approaches (Edwards, 23 Dec 2025, Efremenko et al., 2016).

Related open directions include the search for algebraic-geometric invariants with faster growth rates capable of breaking the quadratic barrier and exploiting structural insights from minimal free resolutions and higher-order relations in Jacobian ideals.